1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2004 Poul-Henning Kamp 5 * Copyright (c) 1994,1997 John S. Dyson 6 * Copyright (c) 2013 The FreeBSD Foundation 7 * All rights reserved. 8 * 9 * Portions of this software were developed by Konstantin Belousov 10 * under sponsorship from the FreeBSD Foundation. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 /* 35 * this file contains a new buffer I/O scheme implementing a coherent 36 * VM object and buffer cache scheme. Pains have been taken to make 37 * sure that the performance degradation associated with schemes such 38 * as this is not realized. 39 * 40 * Author: John S. Dyson 41 * Significant help during the development and debugging phases 42 * had been provided by David Greenman, also of the FreeBSD core team. 43 * 44 * see man buf(9) for more info. 45 */ 46 47 #include <sys/cdefs.h> 48 __FBSDID("$FreeBSD$"); 49 50 #include <sys/param.h> 51 #include <sys/systm.h> 52 #include <sys/bio.h> 53 #include <sys/bitset.h> 54 #include <sys/conf.h> 55 #include <sys/counter.h> 56 #include <sys/buf.h> 57 #include <sys/devicestat.h> 58 #include <sys/eventhandler.h> 59 #include <sys/fail.h> 60 #include <sys/limits.h> 61 #include <sys/lock.h> 62 #include <sys/malloc.h> 63 #include <sys/mount.h> 64 #include <sys/mutex.h> 65 #include <sys/kernel.h> 66 #include <sys/kthread.h> 67 #include <sys/proc.h> 68 #include <sys/racct.h> 69 #include <sys/resourcevar.h> 70 #include <sys/rwlock.h> 71 #include <sys/smp.h> 72 #include <sys/sysctl.h> 73 #include <sys/sysproto.h> 74 #include <sys/vmem.h> 75 #include <sys/vmmeter.h> 76 #include <sys/vnode.h> 77 #include <sys/watchdog.h> 78 #include <geom/geom.h> 79 #include <vm/vm.h> 80 #include <vm/vm_param.h> 81 #include <vm/vm_kern.h> 82 #include <vm/vm_object.h> 83 #include <vm/vm_page.h> 84 #include <vm/vm_pageout.h> 85 #include <vm/vm_pager.h> 86 #include <vm/vm_extern.h> 87 #include <vm/vm_map.h> 88 #include <vm/swap_pager.h> 89 #include "opt_swap.h" 90 91 static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); 92 93 struct bio_ops bioops; /* I/O operation notification */ 94 95 struct buf_ops buf_ops_bio = { 96 .bop_name = "buf_ops_bio", 97 .bop_write = bufwrite, 98 .bop_strategy = bufstrategy, 99 .bop_sync = bufsync, 100 .bop_bdflush = bufbdflush, 101 }; 102 103 struct bufqueue { 104 struct mtx_padalign bq_lock; 105 TAILQ_HEAD(, buf) bq_queue; 106 uint8_t bq_index; 107 uint16_t bq_subqueue; 108 int bq_len; 109 } __aligned(CACHE_LINE_SIZE); 110 111 #define BQ_LOCKPTR(bq) (&(bq)->bq_lock) 112 #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq))) 113 #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq))) 114 #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED) 115 116 struct bufdomain { 117 struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */ 118 struct bufqueue bd_dirtyq; 119 struct bufqueue *bd_cleanq; 120 struct mtx_padalign bd_run_lock; 121 /* Constants */ 122 long bd_maxbufspace; 123 long bd_hibufspace; 124 long bd_lobufspace; 125 long bd_bufspacethresh; 126 int bd_hifreebuffers; 127 int bd_lofreebuffers; 128 int bd_hidirtybuffers; 129 int bd_lodirtybuffers; 130 int bd_dirtybufthresh; 131 int bd_lim; 132 /* atomics */ 133 int bd_wanted; 134 int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers; 135 int __aligned(CACHE_LINE_SIZE) bd_running; 136 long __aligned(CACHE_LINE_SIZE) bd_bufspace; 137 int __aligned(CACHE_LINE_SIZE) bd_freebuffers; 138 } __aligned(CACHE_LINE_SIZE); 139 140 #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock) 141 #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd))) 142 #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd))) 143 #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED) 144 #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock) 145 #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd))) 146 #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd))) 147 #define BD_DOMAIN(bd) (bd - bdomain) 148 149 static struct buf *buf; /* buffer header pool */ 150 extern struct buf *swbuf; /* Swap buffer header pool. */ 151 caddr_t unmapped_buf; 152 153 /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ 154 struct proc *bufdaemonproc; 155 156 static int inmem(struct vnode *vp, daddr_t blkno); 157 static void vm_hold_free_pages(struct buf *bp, int newbsize); 158 static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, 159 vm_offset_t to); 160 static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); 161 static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, 162 vm_page_t m); 163 static void vfs_clean_pages_dirty_buf(struct buf *bp); 164 static void vfs_setdirty_locked_object(struct buf *bp); 165 static void vfs_vmio_invalidate(struct buf *bp); 166 static void vfs_vmio_truncate(struct buf *bp, int npages); 167 static void vfs_vmio_extend(struct buf *bp, int npages, int size); 168 static int vfs_bio_clcheck(struct vnode *vp, int size, 169 daddr_t lblkno, daddr_t blkno); 170 static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, 171 void (*)(struct buf *)); 172 static int buf_flush(struct vnode *vp, struct bufdomain *, int); 173 static int flushbufqueues(struct vnode *, struct bufdomain *, int, int); 174 static void buf_daemon(void); 175 static __inline void bd_wakeup(void); 176 static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); 177 static void bufkva_reclaim(vmem_t *, int); 178 static void bufkva_free(struct buf *); 179 static int buf_import(void *, void **, int, int, int); 180 static void buf_release(void *, void **, int); 181 static void maxbcachebuf_adjust(void); 182 static inline struct bufdomain *bufdomain(struct buf *); 183 static void bq_remove(struct bufqueue *bq, struct buf *bp); 184 static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock); 185 static int buf_recycle(struct bufdomain *, bool kva); 186 static void bq_init(struct bufqueue *bq, int qindex, int cpu, 187 const char *lockname); 188 static void bd_init(struct bufdomain *bd); 189 static int bd_flushall(struct bufdomain *bd); 190 static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS); 191 static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS); 192 193 static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); 194 int vmiodirenable = TRUE; 195 SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, 196 "Use the VM system for directory writes"); 197 long runningbufspace; 198 SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, 199 "Amount of presently outstanding async buffer io"); 200 SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, 201 NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers"); 202 static counter_u64_t bufkvaspace; 203 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, 204 "Kernel virtual memory used for buffers"); 205 static long maxbufspace; 206 SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace, 207 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace, 208 __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L", 209 "Maximum allowed value of bufspace (including metadata)"); 210 static long bufmallocspace; 211 SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, 212 "Amount of malloced memory for buffers"); 213 static long maxbufmallocspace; 214 SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 215 0, "Maximum amount of malloced memory for buffers"); 216 static long lobufspace; 217 SYSCTL_PROC(_vfs, OID_AUTO, lobufspace, 218 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace, 219 __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L", 220 "Minimum amount of buffers we want to have"); 221 long hibufspace; 222 SYSCTL_PROC(_vfs, OID_AUTO, hibufspace, 223 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace, 224 __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L", 225 "Maximum allowed value of bufspace (excluding metadata)"); 226 long bufspacethresh; 227 SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh, 228 CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh, 229 __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L", 230 "Bufspace consumed before waking the daemon to free some"); 231 static counter_u64_t buffreekvacnt; 232 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, 233 "Number of times we have freed the KVA space from some buffer"); 234 static counter_u64_t bufdefragcnt; 235 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, 236 "Number of times we have had to repeat buffer allocation to defragment"); 237 static long lorunningspace; 238 SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 239 CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", 240 "Minimum preferred space used for in-progress I/O"); 241 static long hirunningspace; 242 SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | 243 CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", 244 "Maximum amount of space to use for in-progress I/O"); 245 int dirtybufferflushes; 246 SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 247 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); 248 int bdwriteskip; 249 SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 250 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); 251 int altbufferflushes; 252 SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW, &altbufferflushes, 253 0, "Number of fsync flushes to limit dirty buffers"); 254 static int recursiveflushes; 255 SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW, &recursiveflushes, 256 0, "Number of flushes skipped due to being recursive"); 257 static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS); 258 SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers, 259 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I", 260 "Number of buffers that are dirty (has unwritten changes) at the moment"); 261 static int lodirtybuffers; 262 SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers, 263 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers, 264 __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I", 265 "How many buffers we want to have free before bufdaemon can sleep"); 266 static int hidirtybuffers; 267 SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers, 268 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers, 269 __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I", 270 "When the number of dirty buffers is considered severe"); 271 int dirtybufthresh; 272 SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh, 273 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh, 274 __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I", 275 "Number of bdwrite to bawrite conversions to clear dirty buffers"); 276 static int numfreebuffers; 277 SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, 278 "Number of free buffers"); 279 static int lofreebuffers; 280 SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers, 281 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers, 282 __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I", 283 "Target number of free buffers"); 284 static int hifreebuffers; 285 SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers, 286 CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers, 287 __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I", 288 "Threshold for clean buffer recycling"); 289 static counter_u64_t getnewbufcalls; 290 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, 291 &getnewbufcalls, "Number of calls to getnewbuf"); 292 static counter_u64_t getnewbufrestarts; 293 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, 294 &getnewbufrestarts, 295 "Number of times getnewbuf has had to restart a buffer acquisition"); 296 static counter_u64_t mappingrestarts; 297 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD, 298 &mappingrestarts, 299 "Number of times getblk has had to restart a buffer mapping for " 300 "unmapped buffer"); 301 static counter_u64_t numbufallocfails; 302 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, 303 &numbufallocfails, "Number of times buffer allocations failed"); 304 static int flushbufqtarget = 100; 305 SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, 306 "Amount of work to do in flushbufqueues when helping bufdaemon"); 307 static counter_u64_t notbufdflushes; 308 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, 309 "Number of dirty buffer flushes done by the bufdaemon helpers"); 310 static long barrierwrites; 311 SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW, &barrierwrites, 0, 312 "Number of barrier writes"); 313 SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, 314 &unmapped_buf_allowed, 0, 315 "Permit the use of the unmapped i/o"); 316 int maxbcachebuf = MAXBCACHEBUF; 317 SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, 318 "Maximum size of a buffer cache block"); 319 320 /* 321 * This lock synchronizes access to bd_request. 322 */ 323 static struct mtx_padalign __exclusive_cache_line bdlock; 324 325 /* 326 * This lock protects the runningbufreq and synchronizes runningbufwakeup and 327 * waitrunningbufspace(). 328 */ 329 static struct mtx_padalign __exclusive_cache_line rbreqlock; 330 331 /* 332 * Lock that protects bdirtywait. 333 */ 334 static struct mtx_padalign __exclusive_cache_line bdirtylock; 335 336 /* 337 * Wakeup point for bufdaemon, as well as indicator of whether it is already 338 * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it 339 * is idling. 340 */ 341 static int bd_request; 342 343 /* 344 * Request for the buf daemon to write more buffers than is indicated by 345 * lodirtybuf. This may be necessary to push out excess dependencies or 346 * defragment the address space where a simple count of the number of dirty 347 * buffers is insufficient to characterize the demand for flushing them. 348 */ 349 static int bd_speedupreq; 350 351 /* 352 * Synchronization (sleep/wakeup) variable for active buffer space requests. 353 * Set when wait starts, cleared prior to wakeup(). 354 * Used in runningbufwakeup() and waitrunningbufspace(). 355 */ 356 static int runningbufreq; 357 358 /* 359 * Synchronization for bwillwrite() waiters. 360 */ 361 static int bdirtywait; 362 363 /* 364 * Definitions for the buffer free lists. 365 */ 366 #define QUEUE_NONE 0 /* on no queue */ 367 #define QUEUE_EMPTY 1 /* empty buffer headers */ 368 #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ 369 #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ 370 #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */ 371 372 /* Maximum number of buffer domains. */ 373 #define BUF_DOMAINS 8 374 375 struct bufdomainset bdlodirty; /* Domains > lodirty */ 376 struct bufdomainset bdhidirty; /* Domains > hidirty */ 377 378 /* Configured number of clean queues. */ 379 static int __read_mostly buf_domains; 380 381 BITSET_DEFINE(bufdomainset, BUF_DOMAINS); 382 struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS]; 383 struct bufqueue __exclusive_cache_line bqempty; 384 385 /* 386 * per-cpu empty buffer cache. 387 */ 388 uma_zone_t buf_zone; 389 390 /* 391 * Single global constant for BUF_WMESG, to avoid getting multiple references. 392 * buf_wmesg is referred from macros. 393 */ 394 const char *buf_wmesg = BUF_WMESG; 395 396 static int 397 sysctl_runningspace(SYSCTL_HANDLER_ARGS) 398 { 399 long value; 400 int error; 401 402 value = *(long *)arg1; 403 error = sysctl_handle_long(oidp, &value, 0, req); 404 if (error != 0 || req->newptr == NULL) 405 return (error); 406 mtx_lock(&rbreqlock); 407 if (arg1 == &hirunningspace) { 408 if (value < lorunningspace) 409 error = EINVAL; 410 else 411 hirunningspace = value; 412 } else { 413 KASSERT(arg1 == &lorunningspace, 414 ("%s: unknown arg1", __func__)); 415 if (value > hirunningspace) 416 error = EINVAL; 417 else 418 lorunningspace = value; 419 } 420 mtx_unlock(&rbreqlock); 421 return (error); 422 } 423 424 static int 425 sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS) 426 { 427 int error; 428 int value; 429 int i; 430 431 value = *(int *)arg1; 432 error = sysctl_handle_int(oidp, &value, 0, req); 433 if (error != 0 || req->newptr == NULL) 434 return (error); 435 *(int *)arg1 = value; 436 for (i = 0; i < buf_domains; i++) 437 *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 438 value / buf_domains; 439 440 return (error); 441 } 442 443 static int 444 sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS) 445 { 446 long value; 447 int error; 448 int i; 449 450 value = *(long *)arg1; 451 error = sysctl_handle_long(oidp, &value, 0, req); 452 if (error != 0 || req->newptr == NULL) 453 return (error); 454 *(long *)arg1 = value; 455 for (i = 0; i < buf_domains; i++) 456 *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = 457 value / buf_domains; 458 459 return (error); 460 } 461 462 #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ 463 defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) 464 static int 465 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 466 { 467 long lvalue; 468 int ivalue; 469 int i; 470 471 lvalue = 0; 472 for (i = 0; i < buf_domains; i++) 473 lvalue += bdomain[i].bd_bufspace; 474 if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) 475 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 476 if (lvalue > INT_MAX) 477 /* On overflow, still write out a long to trigger ENOMEM. */ 478 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 479 ivalue = lvalue; 480 return (sysctl_handle_int(oidp, &ivalue, 0, req)); 481 } 482 #else 483 static int 484 sysctl_bufspace(SYSCTL_HANDLER_ARGS) 485 { 486 long lvalue; 487 int i; 488 489 lvalue = 0; 490 for (i = 0; i < buf_domains; i++) 491 lvalue += bdomain[i].bd_bufspace; 492 return (sysctl_handle_long(oidp, &lvalue, 0, req)); 493 } 494 #endif 495 496 static int 497 sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS) 498 { 499 int value; 500 int i; 501 502 value = 0; 503 for (i = 0; i < buf_domains; i++) 504 value += bdomain[i].bd_numdirtybuffers; 505 return (sysctl_handle_int(oidp, &value, 0, req)); 506 } 507 508 /* 509 * bdirtywakeup: 510 * 511 * Wakeup any bwillwrite() waiters. 512 */ 513 static void 514 bdirtywakeup(void) 515 { 516 mtx_lock(&bdirtylock); 517 if (bdirtywait) { 518 bdirtywait = 0; 519 wakeup(&bdirtywait); 520 } 521 mtx_unlock(&bdirtylock); 522 } 523 524 /* 525 * bd_clear: 526 * 527 * Clear a domain from the appropriate bitsets when dirtybuffers 528 * is decremented. 529 */ 530 static void 531 bd_clear(struct bufdomain *bd) 532 { 533 534 mtx_lock(&bdirtylock); 535 if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers) 536 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 537 if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers) 538 BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 539 mtx_unlock(&bdirtylock); 540 } 541 542 /* 543 * bd_set: 544 * 545 * Set a domain in the appropriate bitsets when dirtybuffers 546 * is incremented. 547 */ 548 static void 549 bd_set(struct bufdomain *bd) 550 { 551 552 mtx_lock(&bdirtylock); 553 if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers) 554 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); 555 if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers) 556 BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); 557 mtx_unlock(&bdirtylock); 558 } 559 560 /* 561 * bdirtysub: 562 * 563 * Decrement the numdirtybuffers count by one and wakeup any 564 * threads blocked in bwillwrite(). 565 */ 566 static void 567 bdirtysub(struct buf *bp) 568 { 569 struct bufdomain *bd; 570 int num; 571 572 bd = bufdomain(bp); 573 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1); 574 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 575 bdirtywakeup(); 576 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 577 bd_clear(bd); 578 } 579 580 /* 581 * bdirtyadd: 582 * 583 * Increment the numdirtybuffers count by one and wakeup the buf 584 * daemon if needed. 585 */ 586 static void 587 bdirtyadd(struct buf *bp) 588 { 589 struct bufdomain *bd; 590 int num; 591 592 /* 593 * Only do the wakeup once as we cross the boundary. The 594 * buf daemon will keep running until the condition clears. 595 */ 596 bd = bufdomain(bp); 597 num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1); 598 if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) 599 bd_wakeup(); 600 if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) 601 bd_set(bd); 602 } 603 604 /* 605 * bufspace_daemon_wakeup: 606 * 607 * Wakeup the daemons responsible for freeing clean bufs. 608 */ 609 static void 610 bufspace_daemon_wakeup(struct bufdomain *bd) 611 { 612 613 /* 614 * avoid the lock if the daemon is running. 615 */ 616 if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) { 617 BD_RUN_LOCK(bd); 618 atomic_store_int(&bd->bd_running, 1); 619 wakeup(&bd->bd_running); 620 BD_RUN_UNLOCK(bd); 621 } 622 } 623 624 /* 625 * bufspace_daemon_wait: 626 * 627 * Sleep until the domain falls below a limit or one second passes. 628 */ 629 static void 630 bufspace_daemon_wait(struct bufdomain *bd) 631 { 632 /* 633 * Re-check our limits and sleep. bd_running must be 634 * cleared prior to checking the limits to avoid missed 635 * wakeups. The waker will adjust one of bufspace or 636 * freebuffers prior to checking bd_running. 637 */ 638 BD_RUN_LOCK(bd); 639 atomic_store_int(&bd->bd_running, 0); 640 if (bd->bd_bufspace < bd->bd_bufspacethresh && 641 bd->bd_freebuffers > bd->bd_lofreebuffers) { 642 msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP, 643 "-", hz); 644 } else { 645 /* Avoid spurious wakeups while running. */ 646 atomic_store_int(&bd->bd_running, 1); 647 BD_RUN_UNLOCK(bd); 648 } 649 } 650 651 /* 652 * bufspace_adjust: 653 * 654 * Adjust the reported bufspace for a KVA managed buffer, possibly 655 * waking any waiters. 656 */ 657 static void 658 bufspace_adjust(struct buf *bp, int bufsize) 659 { 660 struct bufdomain *bd; 661 long space; 662 int diff; 663 664 KASSERT((bp->b_flags & B_MALLOC) == 0, 665 ("bufspace_adjust: malloc buf %p", bp)); 666 bd = bufdomain(bp); 667 diff = bufsize - bp->b_bufsize; 668 if (diff < 0) { 669 atomic_subtract_long(&bd->bd_bufspace, -diff); 670 } else if (diff > 0) { 671 space = atomic_fetchadd_long(&bd->bd_bufspace, diff); 672 /* Wake up the daemon on the transition. */ 673 if (space < bd->bd_bufspacethresh && 674 space + diff >= bd->bd_bufspacethresh) 675 bufspace_daemon_wakeup(bd); 676 } 677 bp->b_bufsize = bufsize; 678 } 679 680 /* 681 * bufspace_reserve: 682 * 683 * Reserve bufspace before calling allocbuf(). metadata has a 684 * different space limit than data. 685 */ 686 static int 687 bufspace_reserve(struct bufdomain *bd, int size, bool metadata) 688 { 689 long limit, new; 690 long space; 691 692 if (metadata) 693 limit = bd->bd_maxbufspace; 694 else 695 limit = bd->bd_hibufspace; 696 space = atomic_fetchadd_long(&bd->bd_bufspace, size); 697 new = space + size; 698 if (new > limit) { 699 atomic_subtract_long(&bd->bd_bufspace, size); 700 return (ENOSPC); 701 } 702 703 /* Wake up the daemon on the transition. */ 704 if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh) 705 bufspace_daemon_wakeup(bd); 706 707 return (0); 708 } 709 710 /* 711 * bufspace_release: 712 * 713 * Release reserved bufspace after bufspace_adjust() has consumed it. 714 */ 715 static void 716 bufspace_release(struct bufdomain *bd, int size) 717 { 718 719 atomic_subtract_long(&bd->bd_bufspace, size); 720 } 721 722 /* 723 * bufspace_wait: 724 * 725 * Wait for bufspace, acting as the buf daemon if a locked vnode is 726 * supplied. bd_wanted must be set prior to polling for space. The 727 * operation must be re-tried on return. 728 */ 729 static void 730 bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags, 731 int slpflag, int slptimeo) 732 { 733 struct thread *td; 734 int error, fl, norunbuf; 735 736 if ((gbflags & GB_NOWAIT_BD) != 0) 737 return; 738 739 td = curthread; 740 BD_LOCK(bd); 741 while (bd->bd_wanted) { 742 if (vp != NULL && vp->v_type != VCHR && 743 (td->td_pflags & TDP_BUFNEED) == 0) { 744 BD_UNLOCK(bd); 745 /* 746 * getblk() is called with a vnode locked, and 747 * some majority of the dirty buffers may as 748 * well belong to the vnode. Flushing the 749 * buffers there would make a progress that 750 * cannot be achieved by the buf_daemon, that 751 * cannot lock the vnode. 752 */ 753 norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | 754 (td->td_pflags & TDP_NORUNNINGBUF); 755 756 /* 757 * Play bufdaemon. The getnewbuf() function 758 * may be called while the thread owns lock 759 * for another dirty buffer for the same 760 * vnode, which makes it impossible to use 761 * VOP_FSYNC() there, due to the buffer lock 762 * recursion. 763 */ 764 td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; 765 fl = buf_flush(vp, bd, flushbufqtarget); 766 td->td_pflags &= norunbuf; 767 BD_LOCK(bd); 768 if (fl != 0) 769 continue; 770 if (bd->bd_wanted == 0) 771 break; 772 } 773 error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 774 (PRIBIO + 4) | slpflag, "newbuf", slptimeo); 775 if (error != 0) 776 break; 777 } 778 BD_UNLOCK(bd); 779 } 780 781 782 /* 783 * bufspace_daemon: 784 * 785 * buffer space management daemon. Tries to maintain some marginal 786 * amount of free buffer space so that requesting processes neither 787 * block nor work to reclaim buffers. 788 */ 789 static void 790 bufspace_daemon(void *arg) 791 { 792 struct bufdomain *bd; 793 794 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, 795 SHUTDOWN_PRI_LAST + 100); 796 797 bd = arg; 798 for (;;) { 799 kthread_suspend_check(); 800 801 /* 802 * Free buffers from the clean queue until we meet our 803 * targets. 804 * 805 * Theory of operation: The buffer cache is most efficient 806 * when some free buffer headers and space are always 807 * available to getnewbuf(). This daemon attempts to prevent 808 * the excessive blocking and synchronization associated 809 * with shortfall. It goes through three phases according 810 * demand: 811 * 812 * 1) The daemon wakes up voluntarily once per-second 813 * during idle periods when the counters are below 814 * the wakeup thresholds (bufspacethresh, lofreebuffers). 815 * 816 * 2) The daemon wakes up as we cross the thresholds 817 * ahead of any potential blocking. This may bounce 818 * slightly according to the rate of consumption and 819 * release. 820 * 821 * 3) The daemon and consumers are starved for working 822 * clean buffers. This is the 'bufspace' sleep below 823 * which will inefficiently trade bufs with bqrelse 824 * until we return to condition 2. 825 */ 826 while (bd->bd_bufspace > bd->bd_lobufspace || 827 bd->bd_freebuffers < bd->bd_hifreebuffers) { 828 if (buf_recycle(bd, false) != 0) { 829 if (bd_flushall(bd)) 830 continue; 831 /* 832 * Speedup dirty if we've run out of clean 833 * buffers. This is possible in particular 834 * because softdep may held many bufs locked 835 * pending writes to other bufs which are 836 * marked for delayed write, exhausting 837 * clean space until they are written. 838 */ 839 bd_speedup(); 840 BD_LOCK(bd); 841 if (bd->bd_wanted) { 842 msleep(&bd->bd_wanted, BD_LOCKPTR(bd), 843 PRIBIO|PDROP, "bufspace", hz/10); 844 } else 845 BD_UNLOCK(bd); 846 } 847 maybe_yield(); 848 } 849 bufspace_daemon_wait(bd); 850 } 851 } 852 853 /* 854 * bufmallocadjust: 855 * 856 * Adjust the reported bufspace for a malloc managed buffer, possibly 857 * waking any waiters. 858 */ 859 static void 860 bufmallocadjust(struct buf *bp, int bufsize) 861 { 862 int diff; 863 864 KASSERT((bp->b_flags & B_MALLOC) != 0, 865 ("bufmallocadjust: non-malloc buf %p", bp)); 866 diff = bufsize - bp->b_bufsize; 867 if (diff < 0) 868 atomic_subtract_long(&bufmallocspace, -diff); 869 else 870 atomic_add_long(&bufmallocspace, diff); 871 bp->b_bufsize = bufsize; 872 } 873 874 /* 875 * runningwakeup: 876 * 877 * Wake up processes that are waiting on asynchronous writes to fall 878 * below lorunningspace. 879 */ 880 static void 881 runningwakeup(void) 882 { 883 884 mtx_lock(&rbreqlock); 885 if (runningbufreq) { 886 runningbufreq = 0; 887 wakeup(&runningbufreq); 888 } 889 mtx_unlock(&rbreqlock); 890 } 891 892 /* 893 * runningbufwakeup: 894 * 895 * Decrement the outstanding write count according. 896 */ 897 void 898 runningbufwakeup(struct buf *bp) 899 { 900 long space, bspace; 901 902 bspace = bp->b_runningbufspace; 903 if (bspace == 0) 904 return; 905 space = atomic_fetchadd_long(&runningbufspace, -bspace); 906 KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", 907 space, bspace)); 908 bp->b_runningbufspace = 0; 909 /* 910 * Only acquire the lock and wakeup on the transition from exceeding 911 * the threshold to falling below it. 912 */ 913 if (space < lorunningspace) 914 return; 915 if (space - bspace > lorunningspace) 916 return; 917 runningwakeup(); 918 } 919 920 /* 921 * waitrunningbufspace() 922 * 923 * runningbufspace is a measure of the amount of I/O currently 924 * running. This routine is used in async-write situations to 925 * prevent creating huge backups of pending writes to a device. 926 * Only asynchronous writes are governed by this function. 927 * 928 * This does NOT turn an async write into a sync write. It waits 929 * for earlier writes to complete and generally returns before the 930 * caller's write has reached the device. 931 */ 932 void 933 waitrunningbufspace(void) 934 { 935 936 mtx_lock(&rbreqlock); 937 while (runningbufspace > hirunningspace) { 938 runningbufreq = 1; 939 msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); 940 } 941 mtx_unlock(&rbreqlock); 942 } 943 944 945 /* 946 * vfs_buf_test_cache: 947 * 948 * Called when a buffer is extended. This function clears the B_CACHE 949 * bit if the newly extended portion of the buffer does not contain 950 * valid data. 951 */ 952 static __inline void 953 vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, 954 vm_offset_t size, vm_page_t m) 955 { 956 957 VM_OBJECT_ASSERT_LOCKED(m->object); 958 if (bp->b_flags & B_CACHE) { 959 int base = (foff + off) & PAGE_MASK; 960 if (vm_page_is_valid(m, base, size) == 0) 961 bp->b_flags &= ~B_CACHE; 962 } 963 } 964 965 /* Wake up the buffer daemon if necessary */ 966 static void 967 bd_wakeup(void) 968 { 969 970 mtx_lock(&bdlock); 971 if (bd_request == 0) { 972 bd_request = 1; 973 wakeup(&bd_request); 974 } 975 mtx_unlock(&bdlock); 976 } 977 978 /* 979 * Adjust the maxbcachbuf tunable. 980 */ 981 static void 982 maxbcachebuf_adjust(void) 983 { 984 int i; 985 986 /* 987 * maxbcachebuf must be a power of 2 >= MAXBSIZE. 988 */ 989 i = 2; 990 while (i * 2 <= maxbcachebuf) 991 i *= 2; 992 maxbcachebuf = i; 993 if (maxbcachebuf < MAXBSIZE) 994 maxbcachebuf = MAXBSIZE; 995 if (maxbcachebuf > MAXPHYS) 996 maxbcachebuf = MAXPHYS; 997 if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) 998 printf("maxbcachebuf=%d\n", maxbcachebuf); 999 } 1000 1001 /* 1002 * bd_speedup - speedup the buffer cache flushing code 1003 */ 1004 void 1005 bd_speedup(void) 1006 { 1007 int needwake; 1008 1009 mtx_lock(&bdlock); 1010 needwake = 0; 1011 if (bd_speedupreq == 0 || bd_request == 0) 1012 needwake = 1; 1013 bd_speedupreq = 1; 1014 bd_request = 1; 1015 if (needwake) 1016 wakeup(&bd_request); 1017 mtx_unlock(&bdlock); 1018 } 1019 1020 #ifndef NSWBUF_MIN 1021 #define NSWBUF_MIN 16 1022 #endif 1023 1024 #ifdef __i386__ 1025 #define TRANSIENT_DENOM 5 1026 #else 1027 #define TRANSIENT_DENOM 10 1028 #endif 1029 1030 /* 1031 * Calculating buffer cache scaling values and reserve space for buffer 1032 * headers. This is called during low level kernel initialization and 1033 * may be called more then once. We CANNOT write to the memory area 1034 * being reserved at this time. 1035 */ 1036 caddr_t 1037 kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) 1038 { 1039 int tuned_nbuf; 1040 long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; 1041 1042 /* 1043 * physmem_est is in pages. Convert it to kilobytes (assumes 1044 * PAGE_SIZE is >= 1K) 1045 */ 1046 physmem_est = physmem_est * (PAGE_SIZE / 1024); 1047 1048 maxbcachebuf_adjust(); 1049 /* 1050 * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. 1051 * For the first 64MB of ram nominally allocate sufficient buffers to 1052 * cover 1/4 of our ram. Beyond the first 64MB allocate additional 1053 * buffers to cover 1/10 of our ram over 64MB. When auto-sizing 1054 * the buffer cache we limit the eventual kva reservation to 1055 * maxbcache bytes. 1056 * 1057 * factor represents the 1/4 x ram conversion. 1058 */ 1059 if (nbuf == 0) { 1060 int factor = 4 * BKVASIZE / 1024; 1061 1062 nbuf = 50; 1063 if (physmem_est > 4096) 1064 nbuf += min((physmem_est - 4096) / factor, 1065 65536 / factor); 1066 if (physmem_est > 65536) 1067 nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 1068 32 * 1024 * 1024 / (factor * 5)); 1069 1070 if (maxbcache && nbuf > maxbcache / BKVASIZE) 1071 nbuf = maxbcache / BKVASIZE; 1072 tuned_nbuf = 1; 1073 } else 1074 tuned_nbuf = 0; 1075 1076 /* XXX Avoid unsigned long overflows later on with maxbufspace. */ 1077 maxbuf = (LONG_MAX / 3) / BKVASIZE; 1078 if (nbuf > maxbuf) { 1079 if (!tuned_nbuf) 1080 printf("Warning: nbufs lowered from %d to %ld\n", nbuf, 1081 maxbuf); 1082 nbuf = maxbuf; 1083 } 1084 1085 /* 1086 * Ideal allocation size for the transient bio submap is 10% 1087 * of the maximal space buffer map. This roughly corresponds 1088 * to the amount of the buffer mapped for typical UFS load. 1089 * 1090 * Clip the buffer map to reserve space for the transient 1091 * BIOs, if its extent is bigger than 90% (80% on i386) of the 1092 * maximum buffer map extent on the platform. 1093 * 1094 * The fall-back to the maxbuf in case of maxbcache unset, 1095 * allows to not trim the buffer KVA for the architectures 1096 * with ample KVA space. 1097 */ 1098 if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { 1099 maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; 1100 buf_sz = (long)nbuf * BKVASIZE; 1101 if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * 1102 (TRANSIENT_DENOM - 1)) { 1103 /* 1104 * There is more KVA than memory. Do not 1105 * adjust buffer map size, and assign the rest 1106 * of maxbuf to transient map. 1107 */ 1108 biotmap_sz = maxbuf_sz - buf_sz; 1109 } else { 1110 /* 1111 * Buffer map spans all KVA we could afford on 1112 * this platform. Give 10% (20% on i386) of 1113 * the buffer map to the transient bio map. 1114 */ 1115 biotmap_sz = buf_sz / TRANSIENT_DENOM; 1116 buf_sz -= biotmap_sz; 1117 } 1118 if (biotmap_sz / INT_MAX > MAXPHYS) 1119 bio_transient_maxcnt = INT_MAX; 1120 else 1121 bio_transient_maxcnt = biotmap_sz / MAXPHYS; 1122 /* 1123 * Artificially limit to 1024 simultaneous in-flight I/Os 1124 * using the transient mapping. 1125 */ 1126 if (bio_transient_maxcnt > 1024) 1127 bio_transient_maxcnt = 1024; 1128 if (tuned_nbuf) 1129 nbuf = buf_sz / BKVASIZE; 1130 } 1131 1132 /* 1133 * swbufs are used as temporary holders for I/O, such as paging I/O. 1134 * We have no less then 16 and no more then 256. 1135 */ 1136 nswbuf = min(nbuf / 4, 256); 1137 TUNABLE_INT_FETCH("kern.nswbuf", &nswbuf); 1138 if (nswbuf < NSWBUF_MIN) 1139 nswbuf = NSWBUF_MIN; 1140 1141 /* 1142 * Reserve space for the buffer cache buffers 1143 */ 1144 swbuf = (void *)v; 1145 v = (caddr_t)(swbuf + nswbuf); 1146 buf = (void *)v; 1147 v = (caddr_t)(buf + nbuf); 1148 1149 return(v); 1150 } 1151 1152 /* Initialize the buffer subsystem. Called before use of any buffers. */ 1153 void 1154 bufinit(void) 1155 { 1156 struct buf *bp; 1157 int i; 1158 1159 KASSERT(maxbcachebuf >= MAXBSIZE, 1160 ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, 1161 MAXBSIZE)); 1162 bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); 1163 mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); 1164 mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); 1165 mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); 1166 1167 unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); 1168 1169 /* finally, initialize each buffer header and stick on empty q */ 1170 for (i = 0; i < nbuf; i++) { 1171 bp = &buf[i]; 1172 bzero(bp, sizeof *bp); 1173 bp->b_flags = B_INVAL; 1174 bp->b_rcred = NOCRED; 1175 bp->b_wcred = NOCRED; 1176 bp->b_qindex = QUEUE_NONE; 1177 bp->b_domain = -1; 1178 bp->b_subqueue = mp_maxid + 1; 1179 bp->b_xflags = 0; 1180 bp->b_data = bp->b_kvabase = unmapped_buf; 1181 LIST_INIT(&bp->b_dep); 1182 BUF_LOCKINIT(bp); 1183 bq_insert(&bqempty, bp, false); 1184 } 1185 1186 /* 1187 * maxbufspace is the absolute maximum amount of buffer space we are 1188 * allowed to reserve in KVM and in real terms. The absolute maximum 1189 * is nominally used by metadata. hibufspace is the nominal maximum 1190 * used by most other requests. The differential is required to 1191 * ensure that metadata deadlocks don't occur. 1192 * 1193 * maxbufspace is based on BKVASIZE. Allocating buffers larger then 1194 * this may result in KVM fragmentation which is not handled optimally 1195 * by the system. XXX This is less true with vmem. We could use 1196 * PAGE_SIZE. 1197 */ 1198 maxbufspace = (long)nbuf * BKVASIZE; 1199 hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); 1200 lobufspace = (hibufspace / 20) * 19; /* 95% */ 1201 bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; 1202 1203 /* 1204 * Note: The 16 MiB upper limit for hirunningspace was chosen 1205 * arbitrarily and may need further tuning. It corresponds to 1206 * 128 outstanding write IO requests (if IO size is 128 KiB), 1207 * which fits with many RAID controllers' tagged queuing limits. 1208 * The lower 1 MiB limit is the historical upper limit for 1209 * hirunningspace. 1210 */ 1211 hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 1212 16 * 1024 * 1024), 1024 * 1024); 1213 lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); 1214 1215 /* 1216 * Limit the amount of malloc memory since it is wired permanently into 1217 * the kernel space. Even though this is accounted for in the buffer 1218 * allocation, we don't want the malloced region to grow uncontrolled. 1219 * The malloc scheme improves memory utilization significantly on 1220 * average (small) directories. 1221 */ 1222 maxbufmallocspace = hibufspace / 20; 1223 1224 /* 1225 * Reduce the chance of a deadlock occurring by limiting the number 1226 * of delayed-write dirty buffers we allow to stack up. 1227 */ 1228 hidirtybuffers = nbuf / 4 + 20; 1229 dirtybufthresh = hidirtybuffers * 9 / 10; 1230 /* 1231 * To support extreme low-memory systems, make sure hidirtybuffers 1232 * cannot eat up all available buffer space. This occurs when our 1233 * minimum cannot be met. We try to size hidirtybuffers to 3/4 our 1234 * buffer space assuming BKVASIZE'd buffers. 1235 */ 1236 while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { 1237 hidirtybuffers >>= 1; 1238 } 1239 lodirtybuffers = hidirtybuffers / 2; 1240 1241 /* 1242 * lofreebuffers should be sufficient to avoid stalling waiting on 1243 * buf headers under heavy utilization. The bufs in per-cpu caches 1244 * are counted as free but will be unavailable to threads executing 1245 * on other cpus. 1246 * 1247 * hifreebuffers is the free target for the bufspace daemon. This 1248 * should be set appropriately to limit work per-iteration. 1249 */ 1250 lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); 1251 hifreebuffers = (3 * lofreebuffers) / 2; 1252 numfreebuffers = nbuf; 1253 1254 /* Setup the kva and free list allocators. */ 1255 vmem_set_reclaim(buffer_arena, bufkva_reclaim); 1256 buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf), 1257 NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); 1258 1259 /* 1260 * Size the clean queue according to the amount of buffer space. 1261 * One queue per-256mb up to the max. More queues gives better 1262 * concurrency but less accurate LRU. 1263 */ 1264 buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); 1265 for (i = 0 ; i < buf_domains; i++) { 1266 struct bufdomain *bd; 1267 1268 bd = &bdomain[i]; 1269 bd_init(bd); 1270 bd->bd_freebuffers = nbuf / buf_domains; 1271 bd->bd_hifreebuffers = hifreebuffers / buf_domains; 1272 bd->bd_lofreebuffers = lofreebuffers / buf_domains; 1273 bd->bd_bufspace = 0; 1274 bd->bd_maxbufspace = maxbufspace / buf_domains; 1275 bd->bd_hibufspace = hibufspace / buf_domains; 1276 bd->bd_lobufspace = lobufspace / buf_domains; 1277 bd->bd_bufspacethresh = bufspacethresh / buf_domains; 1278 bd->bd_numdirtybuffers = 0; 1279 bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; 1280 bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; 1281 bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; 1282 /* Don't allow more than 2% of bufs in the per-cpu caches. */ 1283 bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; 1284 } 1285 getnewbufcalls = counter_u64_alloc(M_WAITOK); 1286 getnewbufrestarts = counter_u64_alloc(M_WAITOK); 1287 mappingrestarts = counter_u64_alloc(M_WAITOK); 1288 numbufallocfails = counter_u64_alloc(M_WAITOK); 1289 notbufdflushes = counter_u64_alloc(M_WAITOK); 1290 buffreekvacnt = counter_u64_alloc(M_WAITOK); 1291 bufdefragcnt = counter_u64_alloc(M_WAITOK); 1292 bufkvaspace = counter_u64_alloc(M_WAITOK); 1293 } 1294 1295 #ifdef INVARIANTS 1296 static inline void 1297 vfs_buf_check_mapped(struct buf *bp) 1298 { 1299 1300 KASSERT(bp->b_kvabase != unmapped_buf, 1301 ("mapped buf: b_kvabase was not updated %p", bp)); 1302 KASSERT(bp->b_data != unmapped_buf, 1303 ("mapped buf: b_data was not updated %p", bp)); 1304 KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + 1305 MAXPHYS, ("b_data + b_offset unmapped %p", bp)); 1306 } 1307 1308 static inline void 1309 vfs_buf_check_unmapped(struct buf *bp) 1310 { 1311 1312 KASSERT(bp->b_data == unmapped_buf, 1313 ("unmapped buf: corrupted b_data %p", bp)); 1314 } 1315 1316 #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) 1317 #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) 1318 #else 1319 #define BUF_CHECK_MAPPED(bp) do {} while (0) 1320 #define BUF_CHECK_UNMAPPED(bp) do {} while (0) 1321 #endif 1322 1323 static int 1324 isbufbusy(struct buf *bp) 1325 { 1326 if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || 1327 ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) 1328 return (1); 1329 return (0); 1330 } 1331 1332 /* 1333 * Shutdown the system cleanly to prepare for reboot, halt, or power off. 1334 */ 1335 void 1336 bufshutdown(int show_busybufs) 1337 { 1338 static int first_buf_printf = 1; 1339 struct buf *bp; 1340 int iter, nbusy, pbusy; 1341 #ifndef PREEMPTION 1342 int subiter; 1343 #endif 1344 1345 /* 1346 * Sync filesystems for shutdown 1347 */ 1348 wdog_kern_pat(WD_LASTVAL); 1349 sys_sync(curthread, NULL); 1350 1351 /* 1352 * With soft updates, some buffers that are 1353 * written will be remarked as dirty until other 1354 * buffers are written. 1355 */ 1356 for (iter = pbusy = 0; iter < 20; iter++) { 1357 nbusy = 0; 1358 for (bp = &buf[nbuf]; --bp >= buf; ) 1359 if (isbufbusy(bp)) 1360 nbusy++; 1361 if (nbusy == 0) { 1362 if (first_buf_printf) 1363 printf("All buffers synced."); 1364 break; 1365 } 1366 if (first_buf_printf) { 1367 printf("Syncing disks, buffers remaining... "); 1368 first_buf_printf = 0; 1369 } 1370 printf("%d ", nbusy); 1371 if (nbusy < pbusy) 1372 iter = 0; 1373 pbusy = nbusy; 1374 1375 wdog_kern_pat(WD_LASTVAL); 1376 sys_sync(curthread, NULL); 1377 1378 #ifdef PREEMPTION 1379 /* 1380 * Spin for a while to allow interrupt threads to run. 1381 */ 1382 DELAY(50000 * iter); 1383 #else 1384 /* 1385 * Context switch several times to allow interrupt 1386 * threads to run. 1387 */ 1388 for (subiter = 0; subiter < 50 * iter; subiter++) { 1389 thread_lock(curthread); 1390 mi_switch(SW_VOL, NULL); 1391 thread_unlock(curthread); 1392 DELAY(1000); 1393 } 1394 #endif 1395 } 1396 printf("\n"); 1397 /* 1398 * Count only busy local buffers to prevent forcing 1399 * a fsck if we're just a client of a wedged NFS server 1400 */ 1401 nbusy = 0; 1402 for (bp = &buf[nbuf]; --bp >= buf; ) { 1403 if (isbufbusy(bp)) { 1404 #if 0 1405 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ 1406 if (bp->b_dev == NULL) { 1407 TAILQ_REMOVE(&mountlist, 1408 bp->b_vp->v_mount, mnt_list); 1409 continue; 1410 } 1411 #endif 1412 nbusy++; 1413 if (show_busybufs > 0) { 1414 printf( 1415 "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", 1416 nbusy, bp, bp->b_vp, bp->b_flags, 1417 (intmax_t)bp->b_blkno, 1418 (intmax_t)bp->b_lblkno); 1419 BUF_LOCKPRINTINFO(bp); 1420 if (show_busybufs > 1) 1421 vn_printf(bp->b_vp, 1422 "vnode content: "); 1423 } 1424 } 1425 } 1426 if (nbusy) { 1427 /* 1428 * Failed to sync all blocks. Indicate this and don't 1429 * unmount filesystems (thus forcing an fsck on reboot). 1430 */ 1431 printf("Giving up on %d buffers\n", nbusy); 1432 DELAY(5000000); /* 5 seconds */ 1433 } else { 1434 if (!first_buf_printf) 1435 printf("Final sync complete\n"); 1436 /* 1437 * Unmount filesystems 1438 */ 1439 if (panicstr == NULL) 1440 vfs_unmountall(); 1441 } 1442 swapoff_all(); 1443 DELAY(100000); /* wait for console output to finish */ 1444 } 1445 1446 static void 1447 bpmap_qenter(struct buf *bp) 1448 { 1449 1450 BUF_CHECK_MAPPED(bp); 1451 1452 /* 1453 * bp->b_data is relative to bp->b_offset, but 1454 * bp->b_offset may be offset into the first page. 1455 */ 1456 bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); 1457 pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); 1458 bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | 1459 (vm_offset_t)(bp->b_offset & PAGE_MASK)); 1460 } 1461 1462 static inline struct bufdomain * 1463 bufdomain(struct buf *bp) 1464 { 1465 1466 return (&bdomain[bp->b_domain]); 1467 } 1468 1469 static struct bufqueue * 1470 bufqueue(struct buf *bp) 1471 { 1472 1473 switch (bp->b_qindex) { 1474 case QUEUE_NONE: 1475 /* FALLTHROUGH */ 1476 case QUEUE_SENTINEL: 1477 return (NULL); 1478 case QUEUE_EMPTY: 1479 return (&bqempty); 1480 case QUEUE_DIRTY: 1481 return (&bufdomain(bp)->bd_dirtyq); 1482 case QUEUE_CLEAN: 1483 return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); 1484 default: 1485 break; 1486 } 1487 panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); 1488 } 1489 1490 /* 1491 * Return the locked bufqueue that bp is a member of. 1492 */ 1493 static struct bufqueue * 1494 bufqueue_acquire(struct buf *bp) 1495 { 1496 struct bufqueue *bq, *nbq; 1497 1498 /* 1499 * bp can be pushed from a per-cpu queue to the 1500 * cleanq while we're waiting on the lock. Retry 1501 * if the queues don't match. 1502 */ 1503 bq = bufqueue(bp); 1504 BQ_LOCK(bq); 1505 for (;;) { 1506 nbq = bufqueue(bp); 1507 if (bq == nbq) 1508 break; 1509 BQ_UNLOCK(bq); 1510 BQ_LOCK(nbq); 1511 bq = nbq; 1512 } 1513 return (bq); 1514 } 1515 1516 /* 1517 * binsfree: 1518 * 1519 * Insert the buffer into the appropriate free list. Requires a 1520 * locked buffer on entry and buffer is unlocked before return. 1521 */ 1522 static void 1523 binsfree(struct buf *bp, int qindex) 1524 { 1525 struct bufdomain *bd; 1526 struct bufqueue *bq; 1527 1528 KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, 1529 ("binsfree: Invalid qindex %d", qindex)); 1530 BUF_ASSERT_XLOCKED(bp); 1531 1532 /* 1533 * Handle delayed bremfree() processing. 1534 */ 1535 if (bp->b_flags & B_REMFREE) { 1536 if (bp->b_qindex == qindex) { 1537 bp->b_flags |= B_REUSE; 1538 bp->b_flags &= ~B_REMFREE; 1539 BUF_UNLOCK(bp); 1540 return; 1541 } 1542 bq = bufqueue_acquire(bp); 1543 bq_remove(bq, bp); 1544 BQ_UNLOCK(bq); 1545 } 1546 bd = bufdomain(bp); 1547 if (qindex == QUEUE_CLEAN) { 1548 if (bd->bd_lim != 0) 1549 bq = &bd->bd_subq[PCPU_GET(cpuid)]; 1550 else 1551 bq = bd->bd_cleanq; 1552 } else 1553 bq = &bd->bd_dirtyq; 1554 bq_insert(bq, bp, true); 1555 } 1556 1557 /* 1558 * buf_free: 1559 * 1560 * Free a buffer to the buf zone once it no longer has valid contents. 1561 */ 1562 static void 1563 buf_free(struct buf *bp) 1564 { 1565 1566 if (bp->b_flags & B_REMFREE) 1567 bremfreef(bp); 1568 if (bp->b_vflags & BV_BKGRDINPROG) 1569 panic("losing buffer 1"); 1570 if (bp->b_rcred != NOCRED) { 1571 crfree(bp->b_rcred); 1572 bp->b_rcred = NOCRED; 1573 } 1574 if (bp->b_wcred != NOCRED) { 1575 crfree(bp->b_wcred); 1576 bp->b_wcred = NOCRED; 1577 } 1578 if (!LIST_EMPTY(&bp->b_dep)) 1579 buf_deallocate(bp); 1580 bufkva_free(bp); 1581 atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); 1582 BUF_UNLOCK(bp); 1583 uma_zfree(buf_zone, bp); 1584 } 1585 1586 /* 1587 * buf_import: 1588 * 1589 * Import bufs into the uma cache from the buf list. The system still 1590 * expects a static array of bufs and much of the synchronization 1591 * around bufs assumes type stable storage. As a result, UMA is used 1592 * only as a per-cpu cache of bufs still maintained on a global list. 1593 */ 1594 static int 1595 buf_import(void *arg, void **store, int cnt, int domain, int flags) 1596 { 1597 struct buf *bp; 1598 int i; 1599 1600 BQ_LOCK(&bqempty); 1601 for (i = 0; i < cnt; i++) { 1602 bp = TAILQ_FIRST(&bqempty.bq_queue); 1603 if (bp == NULL) 1604 break; 1605 bq_remove(&bqempty, bp); 1606 store[i] = bp; 1607 } 1608 BQ_UNLOCK(&bqempty); 1609 1610 return (i); 1611 } 1612 1613 /* 1614 * buf_release: 1615 * 1616 * Release bufs from the uma cache back to the buffer queues. 1617 */ 1618 static void 1619 buf_release(void *arg, void **store, int cnt) 1620 { 1621 struct bufqueue *bq; 1622 struct buf *bp; 1623 int i; 1624 1625 bq = &bqempty; 1626 BQ_LOCK(bq); 1627 for (i = 0; i < cnt; i++) { 1628 bp = store[i]; 1629 /* Inline bq_insert() to batch locking. */ 1630 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1631 bp->b_flags &= ~(B_AGE | B_REUSE); 1632 bq->bq_len++; 1633 bp->b_qindex = bq->bq_index; 1634 } 1635 BQ_UNLOCK(bq); 1636 } 1637 1638 /* 1639 * buf_alloc: 1640 * 1641 * Allocate an empty buffer header. 1642 */ 1643 static struct buf * 1644 buf_alloc(struct bufdomain *bd) 1645 { 1646 struct buf *bp; 1647 int freebufs; 1648 1649 /* 1650 * We can only run out of bufs in the buf zone if the average buf 1651 * is less than BKVASIZE. In this case the actual wait/block will 1652 * come from buf_reycle() failing to flush one of these small bufs. 1653 */ 1654 bp = NULL; 1655 freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); 1656 if (freebufs > 0) 1657 bp = uma_zalloc(buf_zone, M_NOWAIT); 1658 if (bp == NULL) { 1659 atomic_fetchadd_int(&bd->bd_freebuffers, 1); 1660 bufspace_daemon_wakeup(bd); 1661 counter_u64_add(numbufallocfails, 1); 1662 return (NULL); 1663 } 1664 /* 1665 * Wake-up the bufspace daemon on transition below threshold. 1666 */ 1667 if (freebufs == bd->bd_lofreebuffers) 1668 bufspace_daemon_wakeup(bd); 1669 1670 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1671 panic("getnewbuf_empty: Locked buf %p on free queue.", bp); 1672 1673 KASSERT(bp->b_vp == NULL, 1674 ("bp: %p still has vnode %p.", bp, bp->b_vp)); 1675 KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, 1676 ("invalid buffer %p flags %#x", bp, bp->b_flags)); 1677 KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, 1678 ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); 1679 KASSERT(bp->b_npages == 0, 1680 ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); 1681 KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); 1682 KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); 1683 1684 bp->b_domain = BD_DOMAIN(bd); 1685 bp->b_flags = 0; 1686 bp->b_ioflags = 0; 1687 bp->b_xflags = 0; 1688 bp->b_vflags = 0; 1689 bp->b_vp = NULL; 1690 bp->b_blkno = bp->b_lblkno = 0; 1691 bp->b_offset = NOOFFSET; 1692 bp->b_iodone = 0; 1693 bp->b_error = 0; 1694 bp->b_resid = 0; 1695 bp->b_bcount = 0; 1696 bp->b_npages = 0; 1697 bp->b_dirtyoff = bp->b_dirtyend = 0; 1698 bp->b_bufobj = NULL; 1699 bp->b_data = bp->b_kvabase = unmapped_buf; 1700 bp->b_fsprivate1 = NULL; 1701 bp->b_fsprivate2 = NULL; 1702 bp->b_fsprivate3 = NULL; 1703 LIST_INIT(&bp->b_dep); 1704 1705 return (bp); 1706 } 1707 1708 /* 1709 * buf_recycle: 1710 * 1711 * Free a buffer from the given bufqueue. kva controls whether the 1712 * freed buf must own some kva resources. This is used for 1713 * defragmenting. 1714 */ 1715 static int 1716 buf_recycle(struct bufdomain *bd, bool kva) 1717 { 1718 struct bufqueue *bq; 1719 struct buf *bp, *nbp; 1720 1721 if (kva) 1722 counter_u64_add(bufdefragcnt, 1); 1723 nbp = NULL; 1724 bq = bd->bd_cleanq; 1725 BQ_LOCK(bq); 1726 KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), 1727 ("buf_recycle: Locks don't match")); 1728 nbp = TAILQ_FIRST(&bq->bq_queue); 1729 1730 /* 1731 * Run scan, possibly freeing data and/or kva mappings on the fly 1732 * depending. 1733 */ 1734 while ((bp = nbp) != NULL) { 1735 /* 1736 * Calculate next bp (we can only use it if we do not 1737 * release the bqlock). 1738 */ 1739 nbp = TAILQ_NEXT(bp, b_freelist); 1740 1741 /* 1742 * If we are defragging then we need a buffer with 1743 * some kva to reclaim. 1744 */ 1745 if (kva && bp->b_kvasize == 0) 1746 continue; 1747 1748 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 1749 continue; 1750 1751 /* 1752 * Implement a second chance algorithm for frequently 1753 * accessed buffers. 1754 */ 1755 if ((bp->b_flags & B_REUSE) != 0) { 1756 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1757 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1758 bp->b_flags &= ~B_REUSE; 1759 BUF_UNLOCK(bp); 1760 continue; 1761 } 1762 1763 /* 1764 * Skip buffers with background writes in progress. 1765 */ 1766 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 1767 BUF_UNLOCK(bp); 1768 continue; 1769 } 1770 1771 KASSERT(bp->b_qindex == QUEUE_CLEAN, 1772 ("buf_recycle: inconsistent queue %d bp %p", 1773 bp->b_qindex, bp)); 1774 KASSERT(bp->b_domain == BD_DOMAIN(bd), 1775 ("getnewbuf: queue domain %d doesn't match request %d", 1776 bp->b_domain, (int)BD_DOMAIN(bd))); 1777 /* 1778 * NOTE: nbp is now entirely invalid. We can only restart 1779 * the scan from this point on. 1780 */ 1781 bq_remove(bq, bp); 1782 BQ_UNLOCK(bq); 1783 1784 /* 1785 * Requeue the background write buffer with error and 1786 * restart the scan. 1787 */ 1788 if ((bp->b_vflags & BV_BKGRDERR) != 0) { 1789 bqrelse(bp); 1790 BQ_LOCK(bq); 1791 nbp = TAILQ_FIRST(&bq->bq_queue); 1792 continue; 1793 } 1794 bp->b_flags |= B_INVAL; 1795 brelse(bp); 1796 return (0); 1797 } 1798 bd->bd_wanted = 1; 1799 BQ_UNLOCK(bq); 1800 1801 return (ENOBUFS); 1802 } 1803 1804 /* 1805 * bremfree: 1806 * 1807 * Mark the buffer for removal from the appropriate free list. 1808 * 1809 */ 1810 void 1811 bremfree(struct buf *bp) 1812 { 1813 1814 CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 1815 KASSERT((bp->b_flags & B_REMFREE) == 0, 1816 ("bremfree: buffer %p already marked for delayed removal.", bp)); 1817 KASSERT(bp->b_qindex != QUEUE_NONE, 1818 ("bremfree: buffer %p not on a queue.", bp)); 1819 BUF_ASSERT_XLOCKED(bp); 1820 1821 bp->b_flags |= B_REMFREE; 1822 } 1823 1824 /* 1825 * bremfreef: 1826 * 1827 * Force an immediate removal from a free list. Used only in nfs when 1828 * it abuses the b_freelist pointer. 1829 */ 1830 void 1831 bremfreef(struct buf *bp) 1832 { 1833 struct bufqueue *bq; 1834 1835 bq = bufqueue_acquire(bp); 1836 bq_remove(bq, bp); 1837 BQ_UNLOCK(bq); 1838 } 1839 1840 static void 1841 bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) 1842 { 1843 1844 mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); 1845 TAILQ_INIT(&bq->bq_queue); 1846 bq->bq_len = 0; 1847 bq->bq_index = qindex; 1848 bq->bq_subqueue = subqueue; 1849 } 1850 1851 static void 1852 bd_init(struct bufdomain *bd) 1853 { 1854 int domain; 1855 int i; 1856 1857 domain = bd - bdomain; 1858 bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; 1859 bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); 1860 bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); 1861 for (i = 0; i <= mp_maxid; i++) 1862 bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, 1863 "bufq clean subqueue lock"); 1864 mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); 1865 } 1866 1867 /* 1868 * bq_remove: 1869 * 1870 * Removes a buffer from the free list, must be called with the 1871 * correct qlock held. 1872 */ 1873 static void 1874 bq_remove(struct bufqueue *bq, struct buf *bp) 1875 { 1876 1877 CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", 1878 bp, bp->b_vp, bp->b_flags); 1879 KASSERT(bp->b_qindex != QUEUE_NONE, 1880 ("bq_remove: buffer %p not on a queue.", bp)); 1881 KASSERT(bufqueue(bp) == bq, 1882 ("bq_remove: Remove buffer %p from wrong queue.", bp)); 1883 1884 BQ_ASSERT_LOCKED(bq); 1885 if (bp->b_qindex != QUEUE_EMPTY) { 1886 BUF_ASSERT_XLOCKED(bp); 1887 } 1888 KASSERT(bq->bq_len >= 1, 1889 ("queue %d underflow", bp->b_qindex)); 1890 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1891 bq->bq_len--; 1892 bp->b_qindex = QUEUE_NONE; 1893 bp->b_flags &= ~(B_REMFREE | B_REUSE); 1894 } 1895 1896 static void 1897 bd_flush(struct bufdomain *bd, struct bufqueue *bq) 1898 { 1899 struct buf *bp; 1900 1901 BQ_ASSERT_LOCKED(bq); 1902 if (bq != bd->bd_cleanq) { 1903 BD_LOCK(bd); 1904 while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { 1905 TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); 1906 TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, 1907 b_freelist); 1908 bp->b_subqueue = bd->bd_cleanq->bq_subqueue; 1909 } 1910 bd->bd_cleanq->bq_len += bq->bq_len; 1911 bq->bq_len = 0; 1912 } 1913 if (bd->bd_wanted) { 1914 bd->bd_wanted = 0; 1915 wakeup(&bd->bd_wanted); 1916 } 1917 if (bq != bd->bd_cleanq) 1918 BD_UNLOCK(bd); 1919 } 1920 1921 static int 1922 bd_flushall(struct bufdomain *bd) 1923 { 1924 struct bufqueue *bq; 1925 int flushed; 1926 int i; 1927 1928 if (bd->bd_lim == 0) 1929 return (0); 1930 flushed = 0; 1931 for (i = 0; i <= mp_maxid; i++) { 1932 bq = &bd->bd_subq[i]; 1933 if (bq->bq_len == 0) 1934 continue; 1935 BQ_LOCK(bq); 1936 bd_flush(bd, bq); 1937 BQ_UNLOCK(bq); 1938 flushed++; 1939 } 1940 1941 return (flushed); 1942 } 1943 1944 static void 1945 bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) 1946 { 1947 struct bufdomain *bd; 1948 1949 if (bp->b_qindex != QUEUE_NONE) 1950 panic("bq_insert: free buffer %p onto another queue?", bp); 1951 1952 bd = bufdomain(bp); 1953 if (bp->b_flags & B_AGE) { 1954 /* Place this buf directly on the real queue. */ 1955 if (bq->bq_index == QUEUE_CLEAN) 1956 bq = bd->bd_cleanq; 1957 BQ_LOCK(bq); 1958 TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); 1959 } else { 1960 BQ_LOCK(bq); 1961 TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); 1962 } 1963 bp->b_flags &= ~(B_AGE | B_REUSE); 1964 bq->bq_len++; 1965 bp->b_qindex = bq->bq_index; 1966 bp->b_subqueue = bq->bq_subqueue; 1967 1968 /* 1969 * Unlock before we notify so that we don't wakeup a waiter that 1970 * fails a trylock on the buf and sleeps again. 1971 */ 1972 if (unlock) 1973 BUF_UNLOCK(bp); 1974 1975 if (bp->b_qindex == QUEUE_CLEAN) { 1976 /* 1977 * Flush the per-cpu queue and notify any waiters. 1978 */ 1979 if (bd->bd_wanted || (bq != bd->bd_cleanq && 1980 bq->bq_len >= bd->bd_lim)) 1981 bd_flush(bd, bq); 1982 } 1983 BQ_UNLOCK(bq); 1984 } 1985 1986 /* 1987 * bufkva_free: 1988 * 1989 * Free the kva allocation for a buffer. 1990 * 1991 */ 1992 static void 1993 bufkva_free(struct buf *bp) 1994 { 1995 1996 #ifdef INVARIANTS 1997 if (bp->b_kvasize == 0) { 1998 KASSERT(bp->b_kvabase == unmapped_buf && 1999 bp->b_data == unmapped_buf, 2000 ("Leaked KVA space on %p", bp)); 2001 } else if (buf_mapped(bp)) 2002 BUF_CHECK_MAPPED(bp); 2003 else 2004 BUF_CHECK_UNMAPPED(bp); 2005 #endif 2006 if (bp->b_kvasize == 0) 2007 return; 2008 2009 vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); 2010 counter_u64_add(bufkvaspace, -bp->b_kvasize); 2011 counter_u64_add(buffreekvacnt, 1); 2012 bp->b_data = bp->b_kvabase = unmapped_buf; 2013 bp->b_kvasize = 0; 2014 } 2015 2016 /* 2017 * bufkva_alloc: 2018 * 2019 * Allocate the buffer KVA and set b_kvasize and b_kvabase. 2020 */ 2021 static int 2022 bufkva_alloc(struct buf *bp, int maxsize, int gbflags) 2023 { 2024 vm_offset_t addr; 2025 int error; 2026 2027 KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, 2028 ("Invalid gbflags 0x%x in %s", gbflags, __func__)); 2029 2030 bufkva_free(bp); 2031 2032 addr = 0; 2033 error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); 2034 if (error != 0) { 2035 /* 2036 * Buffer map is too fragmented. Request the caller 2037 * to defragment the map. 2038 */ 2039 return (error); 2040 } 2041 bp->b_kvabase = (caddr_t)addr; 2042 bp->b_kvasize = maxsize; 2043 counter_u64_add(bufkvaspace, bp->b_kvasize); 2044 if ((gbflags & GB_UNMAPPED) != 0) { 2045 bp->b_data = unmapped_buf; 2046 BUF_CHECK_UNMAPPED(bp); 2047 } else { 2048 bp->b_data = bp->b_kvabase; 2049 BUF_CHECK_MAPPED(bp); 2050 } 2051 return (0); 2052 } 2053 2054 /* 2055 * bufkva_reclaim: 2056 * 2057 * Reclaim buffer kva by freeing buffers holding kva. This is a vmem 2058 * callback that fires to avoid returning failure. 2059 */ 2060 static void 2061 bufkva_reclaim(vmem_t *vmem, int flags) 2062 { 2063 bool done; 2064 int q; 2065 int i; 2066 2067 done = false; 2068 for (i = 0; i < 5; i++) { 2069 for (q = 0; q < buf_domains; q++) 2070 if (buf_recycle(&bdomain[q], true) != 0) 2071 done = true; 2072 if (done) 2073 break; 2074 } 2075 return; 2076 } 2077 2078 /* 2079 * Attempt to initiate asynchronous I/O on read-ahead blocks. We must 2080 * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, 2081 * the buffer is valid and we do not have to do anything. 2082 */ 2083 static void 2084 breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, 2085 struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) 2086 { 2087 struct buf *rabp; 2088 int i; 2089 2090 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 2091 if (inmem(vp, *rablkno)) 2092 continue; 2093 rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); 2094 if ((rabp->b_flags & B_CACHE) != 0) { 2095 brelse(rabp); 2096 continue; 2097 } 2098 if (!TD_IS_IDLETHREAD(curthread)) { 2099 #ifdef RACCT 2100 if (racct_enable) { 2101 PROC_LOCK(curproc); 2102 racct_add_buf(curproc, rabp, 0); 2103 PROC_UNLOCK(curproc); 2104 } 2105 #endif /* RACCT */ 2106 curthread->td_ru.ru_inblock++; 2107 } 2108 rabp->b_flags |= B_ASYNC; 2109 rabp->b_flags &= ~B_INVAL; 2110 if ((flags & GB_CKHASH) != 0) { 2111 rabp->b_flags |= B_CKHASH; 2112 rabp->b_ckhashcalc = ckhashfunc; 2113 } 2114 rabp->b_ioflags &= ~BIO_ERROR; 2115 rabp->b_iocmd = BIO_READ; 2116 if (rabp->b_rcred == NOCRED && cred != NOCRED) 2117 rabp->b_rcred = crhold(cred); 2118 vfs_busy_pages(rabp, 0); 2119 BUF_KERNPROC(rabp); 2120 rabp->b_iooffset = dbtob(rabp->b_blkno); 2121 bstrategy(rabp); 2122 } 2123 } 2124 2125 /* 2126 * Entry point for bread() and breadn() via #defines in sys/buf.h. 2127 * 2128 * Get a buffer with the specified data. Look in the cache first. We 2129 * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE 2130 * is set, the buffer is valid and we do not have to do anything, see 2131 * getblk(). Also starts asynchronous I/O on read-ahead blocks. 2132 * 2133 * Always return a NULL buffer pointer (in bpp) when returning an error. 2134 */ 2135 int 2136 breadn_flags(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablkno, 2137 int *rabsize, int cnt, struct ucred *cred, int flags, 2138 void (*ckhashfunc)(struct buf *), struct buf **bpp) 2139 { 2140 struct buf *bp; 2141 int readwait, rv; 2142 2143 CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); 2144 /* 2145 * Can only return NULL if GB_LOCK_NOWAIT flag is specified. 2146 */ 2147 *bpp = bp = getblk(vp, blkno, size, 0, 0, flags); 2148 if (bp == NULL) 2149 return (EBUSY); 2150 2151 /* 2152 * If not found in cache, do some I/O 2153 */ 2154 readwait = 0; 2155 if ((bp->b_flags & B_CACHE) == 0) { 2156 if (!TD_IS_IDLETHREAD(curthread)) { 2157 #ifdef RACCT 2158 if (racct_enable) { 2159 PROC_LOCK(curproc); 2160 racct_add_buf(curproc, bp, 0); 2161 PROC_UNLOCK(curproc); 2162 } 2163 #endif /* RACCT */ 2164 curthread->td_ru.ru_inblock++; 2165 } 2166 bp->b_iocmd = BIO_READ; 2167 bp->b_flags &= ~B_INVAL; 2168 if ((flags & GB_CKHASH) != 0) { 2169 bp->b_flags |= B_CKHASH; 2170 bp->b_ckhashcalc = ckhashfunc; 2171 } 2172 bp->b_ioflags &= ~BIO_ERROR; 2173 if (bp->b_rcred == NOCRED && cred != NOCRED) 2174 bp->b_rcred = crhold(cred); 2175 vfs_busy_pages(bp, 0); 2176 bp->b_iooffset = dbtob(bp->b_blkno); 2177 bstrategy(bp); 2178 ++readwait; 2179 } 2180 2181 /* 2182 * Attempt to initiate asynchronous I/O on read-ahead blocks. 2183 */ 2184 breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); 2185 2186 rv = 0; 2187 if (readwait) { 2188 rv = bufwait(bp); 2189 if (rv != 0) { 2190 brelse(bp); 2191 *bpp = NULL; 2192 } 2193 } 2194 return (rv); 2195 } 2196 2197 /* 2198 * Write, release buffer on completion. (Done by iodone 2199 * if async). Do not bother writing anything if the buffer 2200 * is invalid. 2201 * 2202 * Note that we set B_CACHE here, indicating that buffer is 2203 * fully valid and thus cacheable. This is true even of NFS 2204 * now so we set it generally. This could be set either here 2205 * or in biodone() since the I/O is synchronous. We put it 2206 * here. 2207 */ 2208 int 2209 bufwrite(struct buf *bp) 2210 { 2211 int oldflags; 2212 struct vnode *vp; 2213 long space; 2214 int vp_md; 2215 2216 CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2217 if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { 2218 bp->b_flags |= B_INVAL | B_RELBUF; 2219 bp->b_flags &= ~B_CACHE; 2220 brelse(bp); 2221 return (ENXIO); 2222 } 2223 if (bp->b_flags & B_INVAL) { 2224 brelse(bp); 2225 return (0); 2226 } 2227 2228 if (bp->b_flags & B_BARRIER) 2229 barrierwrites++; 2230 2231 oldflags = bp->b_flags; 2232 2233 BUF_ASSERT_HELD(bp); 2234 2235 KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), 2236 ("FFS background buffer should not get here %p", bp)); 2237 2238 vp = bp->b_vp; 2239 if (vp) 2240 vp_md = vp->v_vflag & VV_MD; 2241 else 2242 vp_md = 0; 2243 2244 /* 2245 * Mark the buffer clean. Increment the bufobj write count 2246 * before bundirty() call, to prevent other thread from seeing 2247 * empty dirty list and zero counter for writes in progress, 2248 * falsely indicating that the bufobj is clean. 2249 */ 2250 bufobj_wref(bp->b_bufobj); 2251 bundirty(bp); 2252 2253 bp->b_flags &= ~B_DONE; 2254 bp->b_ioflags &= ~BIO_ERROR; 2255 bp->b_flags |= B_CACHE; 2256 bp->b_iocmd = BIO_WRITE; 2257 2258 vfs_busy_pages(bp, 1); 2259 2260 /* 2261 * Normal bwrites pipeline writes 2262 */ 2263 bp->b_runningbufspace = bp->b_bufsize; 2264 space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); 2265 2266 if (!TD_IS_IDLETHREAD(curthread)) { 2267 #ifdef RACCT 2268 if (racct_enable) { 2269 PROC_LOCK(curproc); 2270 racct_add_buf(curproc, bp, 1); 2271 PROC_UNLOCK(curproc); 2272 } 2273 #endif /* RACCT */ 2274 curthread->td_ru.ru_oublock++; 2275 } 2276 if (oldflags & B_ASYNC) 2277 BUF_KERNPROC(bp); 2278 bp->b_iooffset = dbtob(bp->b_blkno); 2279 buf_track(bp, __func__); 2280 bstrategy(bp); 2281 2282 if ((oldflags & B_ASYNC) == 0) { 2283 int rtval = bufwait(bp); 2284 brelse(bp); 2285 return (rtval); 2286 } else if (space > hirunningspace) { 2287 /* 2288 * don't allow the async write to saturate the I/O 2289 * system. We will not deadlock here because 2290 * we are blocking waiting for I/O that is already in-progress 2291 * to complete. We do not block here if it is the update 2292 * or syncer daemon trying to clean up as that can lead 2293 * to deadlock. 2294 */ 2295 if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) 2296 waitrunningbufspace(); 2297 } 2298 2299 return (0); 2300 } 2301 2302 void 2303 bufbdflush(struct bufobj *bo, struct buf *bp) 2304 { 2305 struct buf *nbp; 2306 2307 if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { 2308 (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); 2309 altbufferflushes++; 2310 } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { 2311 BO_LOCK(bo); 2312 /* 2313 * Try to find a buffer to flush. 2314 */ 2315 TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { 2316 if ((nbp->b_vflags & BV_BKGRDINPROG) || 2317 BUF_LOCK(nbp, 2318 LK_EXCLUSIVE | LK_NOWAIT, NULL)) 2319 continue; 2320 if (bp == nbp) 2321 panic("bdwrite: found ourselves"); 2322 BO_UNLOCK(bo); 2323 /* Don't countdeps with the bo lock held. */ 2324 if (buf_countdeps(nbp, 0)) { 2325 BO_LOCK(bo); 2326 BUF_UNLOCK(nbp); 2327 continue; 2328 } 2329 if (nbp->b_flags & B_CLUSTEROK) { 2330 vfs_bio_awrite(nbp); 2331 } else { 2332 bremfree(nbp); 2333 bawrite(nbp); 2334 } 2335 dirtybufferflushes++; 2336 break; 2337 } 2338 if (nbp == NULL) 2339 BO_UNLOCK(bo); 2340 } 2341 } 2342 2343 /* 2344 * Delayed write. (Buffer is marked dirty). Do not bother writing 2345 * anything if the buffer is marked invalid. 2346 * 2347 * Note that since the buffer must be completely valid, we can safely 2348 * set B_CACHE. In fact, we have to set B_CACHE here rather then in 2349 * biodone() in order to prevent getblk from writing the buffer 2350 * out synchronously. 2351 */ 2352 void 2353 bdwrite(struct buf *bp) 2354 { 2355 struct thread *td = curthread; 2356 struct vnode *vp; 2357 struct bufobj *bo; 2358 2359 CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2360 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2361 KASSERT((bp->b_flags & B_BARRIER) == 0, 2362 ("Barrier request in delayed write %p", bp)); 2363 BUF_ASSERT_HELD(bp); 2364 2365 if (bp->b_flags & B_INVAL) { 2366 brelse(bp); 2367 return; 2368 } 2369 2370 /* 2371 * If we have too many dirty buffers, don't create any more. 2372 * If we are wildly over our limit, then force a complete 2373 * cleanup. Otherwise, just keep the situation from getting 2374 * out of control. Note that we have to avoid a recursive 2375 * disaster and not try to clean up after our own cleanup! 2376 */ 2377 vp = bp->b_vp; 2378 bo = bp->b_bufobj; 2379 if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { 2380 td->td_pflags |= TDP_INBDFLUSH; 2381 BO_BDFLUSH(bo, bp); 2382 td->td_pflags &= ~TDP_INBDFLUSH; 2383 } else 2384 recursiveflushes++; 2385 2386 bdirty(bp); 2387 /* 2388 * Set B_CACHE, indicating that the buffer is fully valid. This is 2389 * true even of NFS now. 2390 */ 2391 bp->b_flags |= B_CACHE; 2392 2393 /* 2394 * This bmap keeps the system from needing to do the bmap later, 2395 * perhaps when the system is attempting to do a sync. Since it 2396 * is likely that the indirect block -- or whatever other datastructure 2397 * that the filesystem needs is still in memory now, it is a good 2398 * thing to do this. Note also, that if the pageout daemon is 2399 * requesting a sync -- there might not be enough memory to do 2400 * the bmap then... So, this is important to do. 2401 */ 2402 if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { 2403 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 2404 } 2405 2406 buf_track(bp, __func__); 2407 2408 /* 2409 * Set the *dirty* buffer range based upon the VM system dirty 2410 * pages. 2411 * 2412 * Mark the buffer pages as clean. We need to do this here to 2413 * satisfy the vnode_pager and the pageout daemon, so that it 2414 * thinks that the pages have been "cleaned". Note that since 2415 * the pages are in a delayed write buffer -- the VFS layer 2416 * "will" see that the pages get written out on the next sync, 2417 * or perhaps the cluster will be completed. 2418 */ 2419 vfs_clean_pages_dirty_buf(bp); 2420 bqrelse(bp); 2421 2422 /* 2423 * note: we cannot initiate I/O from a bdwrite even if we wanted to, 2424 * due to the softdep code. 2425 */ 2426 } 2427 2428 /* 2429 * bdirty: 2430 * 2431 * Turn buffer into delayed write request. We must clear BIO_READ and 2432 * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to 2433 * itself to properly update it in the dirty/clean lists. We mark it 2434 * B_DONE to ensure that any asynchronization of the buffer properly 2435 * clears B_DONE ( else a panic will occur later ). 2436 * 2437 * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which 2438 * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() 2439 * should only be called if the buffer is known-good. 2440 * 2441 * Since the buffer is not on a queue, we do not update the numfreebuffers 2442 * count. 2443 * 2444 * The buffer must be on QUEUE_NONE. 2445 */ 2446 void 2447 bdirty(struct buf *bp) 2448 { 2449 2450 CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", 2451 bp, bp->b_vp, bp->b_flags); 2452 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2453 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2454 ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2455 BUF_ASSERT_HELD(bp); 2456 bp->b_flags &= ~(B_RELBUF); 2457 bp->b_iocmd = BIO_WRITE; 2458 2459 if ((bp->b_flags & B_DELWRI) == 0) { 2460 bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; 2461 reassignbuf(bp); 2462 bdirtyadd(bp); 2463 } 2464 } 2465 2466 /* 2467 * bundirty: 2468 * 2469 * Clear B_DELWRI for buffer. 2470 * 2471 * Since the buffer is not on a queue, we do not update the numfreebuffers 2472 * count. 2473 * 2474 * The buffer must be on QUEUE_NONE. 2475 */ 2476 2477 void 2478 bundirty(struct buf *bp) 2479 { 2480 2481 CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2482 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); 2483 KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, 2484 ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); 2485 BUF_ASSERT_HELD(bp); 2486 2487 if (bp->b_flags & B_DELWRI) { 2488 bp->b_flags &= ~B_DELWRI; 2489 reassignbuf(bp); 2490 bdirtysub(bp); 2491 } 2492 /* 2493 * Since it is now being written, we can clear its deferred write flag. 2494 */ 2495 bp->b_flags &= ~B_DEFERRED; 2496 } 2497 2498 /* 2499 * bawrite: 2500 * 2501 * Asynchronous write. Start output on a buffer, but do not wait for 2502 * it to complete. The buffer is released when the output completes. 2503 * 2504 * bwrite() ( or the VOP routine anyway ) is responsible for handling 2505 * B_INVAL buffers. Not us. 2506 */ 2507 void 2508 bawrite(struct buf *bp) 2509 { 2510 2511 bp->b_flags |= B_ASYNC; 2512 (void) bwrite(bp); 2513 } 2514 2515 /* 2516 * babarrierwrite: 2517 * 2518 * Asynchronous barrier write. Start output on a buffer, but do not 2519 * wait for it to complete. Place a write barrier after this write so 2520 * that this buffer and all buffers written before it are committed to 2521 * the disk before any buffers written after this write are committed 2522 * to the disk. The buffer is released when the output completes. 2523 */ 2524 void 2525 babarrierwrite(struct buf *bp) 2526 { 2527 2528 bp->b_flags |= B_ASYNC | B_BARRIER; 2529 (void) bwrite(bp); 2530 } 2531 2532 /* 2533 * bbarrierwrite: 2534 * 2535 * Synchronous barrier write. Start output on a buffer and wait for 2536 * it to complete. Place a write barrier after this write so that 2537 * this buffer and all buffers written before it are committed to 2538 * the disk before any buffers written after this write are committed 2539 * to the disk. The buffer is released when the output completes. 2540 */ 2541 int 2542 bbarrierwrite(struct buf *bp) 2543 { 2544 2545 bp->b_flags |= B_BARRIER; 2546 return (bwrite(bp)); 2547 } 2548 2549 /* 2550 * bwillwrite: 2551 * 2552 * Called prior to the locking of any vnodes when we are expecting to 2553 * write. We do not want to starve the buffer cache with too many 2554 * dirty buffers so we block here. By blocking prior to the locking 2555 * of any vnodes we attempt to avoid the situation where a locked vnode 2556 * prevents the various system daemons from flushing related buffers. 2557 */ 2558 void 2559 bwillwrite(void) 2560 { 2561 2562 if (buf_dirty_count_severe()) { 2563 mtx_lock(&bdirtylock); 2564 while (buf_dirty_count_severe()) { 2565 bdirtywait = 1; 2566 msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), 2567 "flswai", 0); 2568 } 2569 mtx_unlock(&bdirtylock); 2570 } 2571 } 2572 2573 /* 2574 * Return true if we have too many dirty buffers. 2575 */ 2576 int 2577 buf_dirty_count_severe(void) 2578 { 2579 2580 return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); 2581 } 2582 2583 /* 2584 * brelse: 2585 * 2586 * Release a busy buffer and, if requested, free its resources. The 2587 * buffer will be stashed in the appropriate bufqueue[] allowing it 2588 * to be accessed later as a cache entity or reused for other purposes. 2589 */ 2590 void 2591 brelse(struct buf *bp) 2592 { 2593 struct mount *v_mnt; 2594 int qindex; 2595 2596 /* 2597 * Many functions erroneously call brelse with a NULL bp under rare 2598 * error conditions. Simply return when called with a NULL bp. 2599 */ 2600 if (bp == NULL) 2601 return; 2602 CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", 2603 bp, bp->b_vp, bp->b_flags); 2604 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2605 ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2606 KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, 2607 ("brelse: non-VMIO buffer marked NOREUSE")); 2608 2609 if (BUF_LOCKRECURSED(bp)) { 2610 /* 2611 * Do not process, in particular, do not handle the 2612 * B_INVAL/B_RELBUF and do not release to free list. 2613 */ 2614 BUF_UNLOCK(bp); 2615 return; 2616 } 2617 2618 if (bp->b_flags & B_MANAGED) { 2619 bqrelse(bp); 2620 return; 2621 } 2622 2623 if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { 2624 BO_LOCK(bp->b_bufobj); 2625 bp->b_vflags &= ~BV_BKGRDERR; 2626 BO_UNLOCK(bp->b_bufobj); 2627 bdirty(bp); 2628 } 2629 if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && 2630 (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && 2631 !(bp->b_flags & B_INVAL)) { 2632 /* 2633 * Failed write, redirty. All errors except ENXIO (which 2634 * means the device is gone) are treated as being 2635 * transient. 2636 * 2637 * XXX Treating EIO as transient is not correct; the 2638 * contract with the local storage device drivers is that 2639 * they will only return EIO once the I/O is no longer 2640 * retriable. Network I/O also respects this through the 2641 * guarantees of TCP and/or the internal retries of NFS. 2642 * ENOMEM might be transient, but we also have no way of 2643 * knowing when its ok to retry/reschedule. In general, 2644 * this entire case should be made obsolete through better 2645 * error handling/recovery and resource scheduling. 2646 * 2647 * Do this also for buffers that failed with ENXIO, but have 2648 * non-empty dependencies - the soft updates code might need 2649 * to access the buffer to untangle them. 2650 * 2651 * Must clear BIO_ERROR to prevent pages from being scrapped. 2652 */ 2653 bp->b_ioflags &= ~BIO_ERROR; 2654 bdirty(bp); 2655 } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || 2656 (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { 2657 /* 2658 * Either a failed read I/O, or we were asked to free or not 2659 * cache the buffer, or we failed to write to a device that's 2660 * no longer present. 2661 */ 2662 bp->b_flags |= B_INVAL; 2663 if (!LIST_EMPTY(&bp->b_dep)) 2664 buf_deallocate(bp); 2665 if (bp->b_flags & B_DELWRI) 2666 bdirtysub(bp); 2667 bp->b_flags &= ~(B_DELWRI | B_CACHE); 2668 if ((bp->b_flags & B_VMIO) == 0) { 2669 allocbuf(bp, 0); 2670 if (bp->b_vp) 2671 brelvp(bp); 2672 } 2673 } 2674 2675 /* 2676 * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() 2677 * is called with B_DELWRI set, the underlying pages may wind up 2678 * getting freed causing a previous write (bdwrite()) to get 'lost' 2679 * because pages associated with a B_DELWRI bp are marked clean. 2680 * 2681 * We still allow the B_INVAL case to call vfs_vmio_truncate(), even 2682 * if B_DELWRI is set. 2683 */ 2684 if (bp->b_flags & B_DELWRI) 2685 bp->b_flags &= ~B_RELBUF; 2686 2687 /* 2688 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 2689 * constituted, not even NFS buffers now. Two flags effect this. If 2690 * B_INVAL, the struct buf is invalidated but the VM object is kept 2691 * around ( i.e. so it is trivial to reconstitute the buffer later ). 2692 * 2693 * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be 2694 * invalidated. BIO_ERROR cannot be set for a failed write unless the 2695 * buffer is also B_INVAL because it hits the re-dirtying code above. 2696 * 2697 * Normally we can do this whether a buffer is B_DELWRI or not. If 2698 * the buffer is an NFS buffer, it is tracking piecemeal writes or 2699 * the commit state and we cannot afford to lose the buffer. If the 2700 * buffer has a background write in progress, we need to keep it 2701 * around to prevent it from being reconstituted and starting a second 2702 * background write. 2703 */ 2704 2705 v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; 2706 2707 if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || 2708 (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && 2709 (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || 2710 vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) { 2711 vfs_vmio_invalidate(bp); 2712 allocbuf(bp, 0); 2713 } 2714 2715 if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || 2716 (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { 2717 allocbuf(bp, 0); 2718 bp->b_flags &= ~B_NOREUSE; 2719 if (bp->b_vp != NULL) 2720 brelvp(bp); 2721 } 2722 2723 /* 2724 * If the buffer has junk contents signal it and eventually 2725 * clean up B_DELWRI and diassociate the vnode so that gbincore() 2726 * doesn't find it. 2727 */ 2728 if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || 2729 (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) 2730 bp->b_flags |= B_INVAL; 2731 if (bp->b_flags & B_INVAL) { 2732 if (bp->b_flags & B_DELWRI) 2733 bundirty(bp); 2734 if (bp->b_vp) 2735 brelvp(bp); 2736 } 2737 2738 buf_track(bp, __func__); 2739 2740 /* buffers with no memory */ 2741 if (bp->b_bufsize == 0) { 2742 buf_free(bp); 2743 return; 2744 } 2745 /* buffers with junk contents */ 2746 if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || 2747 (bp->b_ioflags & BIO_ERROR)) { 2748 bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); 2749 if (bp->b_vflags & BV_BKGRDINPROG) 2750 panic("losing buffer 2"); 2751 qindex = QUEUE_CLEAN; 2752 bp->b_flags |= B_AGE; 2753 /* remaining buffers */ 2754 } else if (bp->b_flags & B_DELWRI) 2755 qindex = QUEUE_DIRTY; 2756 else 2757 qindex = QUEUE_CLEAN; 2758 2759 if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) 2760 panic("brelse: not dirty"); 2761 2762 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); 2763 /* binsfree unlocks bp. */ 2764 binsfree(bp, qindex); 2765 } 2766 2767 /* 2768 * Release a buffer back to the appropriate queue but do not try to free 2769 * it. The buffer is expected to be used again soon. 2770 * 2771 * bqrelse() is used by bdwrite() to requeue a delayed write, and used by 2772 * biodone() to requeue an async I/O on completion. It is also used when 2773 * known good buffers need to be requeued but we think we may need the data 2774 * again soon. 2775 * 2776 * XXX we should be able to leave the B_RELBUF hint set on completion. 2777 */ 2778 void 2779 bqrelse(struct buf *bp) 2780 { 2781 int qindex; 2782 2783 CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 2784 KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), 2785 ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); 2786 2787 qindex = QUEUE_NONE; 2788 if (BUF_LOCKRECURSED(bp)) { 2789 /* do not release to free list */ 2790 BUF_UNLOCK(bp); 2791 return; 2792 } 2793 bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 2794 2795 if (bp->b_flags & B_MANAGED) { 2796 if (bp->b_flags & B_REMFREE) 2797 bremfreef(bp); 2798 goto out; 2799 } 2800 2801 /* buffers with stale but valid contents */ 2802 if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | 2803 BV_BKGRDERR)) == BV_BKGRDERR) { 2804 BO_LOCK(bp->b_bufobj); 2805 bp->b_vflags &= ~BV_BKGRDERR; 2806 BO_UNLOCK(bp->b_bufobj); 2807 qindex = QUEUE_DIRTY; 2808 } else { 2809 if ((bp->b_flags & B_DELWRI) == 0 && 2810 (bp->b_xflags & BX_VNDIRTY)) 2811 panic("bqrelse: not dirty"); 2812 if ((bp->b_flags & B_NOREUSE) != 0) { 2813 brelse(bp); 2814 return; 2815 } 2816 qindex = QUEUE_CLEAN; 2817 } 2818 buf_track(bp, __func__); 2819 /* binsfree unlocks bp. */ 2820 binsfree(bp, qindex); 2821 return; 2822 2823 out: 2824 buf_track(bp, __func__); 2825 /* unlock */ 2826 BUF_UNLOCK(bp); 2827 } 2828 2829 /* 2830 * Complete I/O to a VMIO backed page. Validate the pages as appropriate, 2831 * restore bogus pages. 2832 */ 2833 static void 2834 vfs_vmio_iodone(struct buf *bp) 2835 { 2836 vm_ooffset_t foff; 2837 vm_page_t m; 2838 vm_object_t obj; 2839 struct vnode *vp; 2840 int i, iosize, resid; 2841 bool bogus; 2842 2843 obj = bp->b_bufobj->bo_object; 2844 KASSERT(obj->paging_in_progress >= bp->b_npages, 2845 ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", 2846 obj->paging_in_progress, bp->b_npages)); 2847 2848 vp = bp->b_vp; 2849 KASSERT(vp->v_holdcnt > 0, 2850 ("vfs_vmio_iodone: vnode %p has zero hold count", vp)); 2851 KASSERT(vp->v_object != NULL, 2852 ("vfs_vmio_iodone: vnode %p has no vm_object", vp)); 2853 2854 foff = bp->b_offset; 2855 KASSERT(bp->b_offset != NOOFFSET, 2856 ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); 2857 2858 bogus = false; 2859 iosize = bp->b_bcount - bp->b_resid; 2860 VM_OBJECT_WLOCK(obj); 2861 for (i = 0; i < bp->b_npages; i++) { 2862 resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; 2863 if (resid > iosize) 2864 resid = iosize; 2865 2866 /* 2867 * cleanup bogus pages, restoring the originals 2868 */ 2869 m = bp->b_pages[i]; 2870 if (m == bogus_page) { 2871 bogus = true; 2872 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 2873 if (m == NULL) 2874 panic("biodone: page disappeared!"); 2875 bp->b_pages[i] = m; 2876 } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { 2877 /* 2878 * In the write case, the valid and clean bits are 2879 * already changed correctly ( see bdwrite() ), so we 2880 * only need to do this here in the read case. 2881 */ 2882 KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, 2883 resid)) == 0, ("vfs_vmio_iodone: page %p " 2884 "has unexpected dirty bits", m)); 2885 vfs_page_set_valid(bp, foff, m); 2886 } 2887 KASSERT(OFF_TO_IDX(foff) == m->pindex, 2888 ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", 2889 (intmax_t)foff, (uintmax_t)m->pindex)); 2890 2891 vm_page_sunbusy(m); 2892 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 2893 iosize -= resid; 2894 } 2895 vm_object_pip_wakeupn(obj, bp->b_npages); 2896 VM_OBJECT_WUNLOCK(obj); 2897 if (bogus && buf_mapped(bp)) { 2898 BUF_CHECK_MAPPED(bp); 2899 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 2900 bp->b_pages, bp->b_npages); 2901 } 2902 } 2903 2904 /* 2905 * Unwire a page held by a buf and either free it or update the page queues to 2906 * reflect its recent use. 2907 */ 2908 static void 2909 vfs_vmio_unwire(struct buf *bp, vm_page_t m) 2910 { 2911 bool freed; 2912 2913 vm_page_lock(m); 2914 if (vm_page_unwire_noq(m)) { 2915 if ((bp->b_flags & B_DIRECT) != 0) 2916 freed = vm_page_try_to_free(m); 2917 else 2918 freed = false; 2919 if (!freed) { 2920 /* 2921 * Use a racy check of the valid bits to determine 2922 * whether we can accelerate reclamation of the page. 2923 * The valid bits will be stable unless the page is 2924 * being mapped or is referenced by multiple buffers, 2925 * and in those cases we expect races to be rare. At 2926 * worst we will either accelerate reclamation of a 2927 * valid page and violate LRU, or unnecessarily defer 2928 * reclamation of an invalid page. 2929 * 2930 * The B_NOREUSE flag marks data that is not expected to 2931 * be reused, so accelerate reclamation in that case 2932 * too. Otherwise, maintain LRU. 2933 */ 2934 if (m->valid == 0 || (bp->b_flags & B_NOREUSE) != 0) 2935 vm_page_deactivate_noreuse(m); 2936 else if (vm_page_active(m)) 2937 vm_page_reference(m); 2938 else 2939 vm_page_deactivate(m); 2940 } 2941 } 2942 vm_page_unlock(m); 2943 } 2944 2945 /* 2946 * Perform page invalidation when a buffer is released. The fully invalid 2947 * pages will be reclaimed later in vfs_vmio_truncate(). 2948 */ 2949 static void 2950 vfs_vmio_invalidate(struct buf *bp) 2951 { 2952 vm_object_t obj; 2953 vm_page_t m; 2954 int i, resid, poffset, presid; 2955 2956 if (buf_mapped(bp)) { 2957 BUF_CHECK_MAPPED(bp); 2958 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); 2959 } else 2960 BUF_CHECK_UNMAPPED(bp); 2961 /* 2962 * Get the base offset and length of the buffer. Note that 2963 * in the VMIO case if the buffer block size is not 2964 * page-aligned then b_data pointer may not be page-aligned. 2965 * But our b_pages[] array *IS* page aligned. 2966 * 2967 * block sizes less then DEV_BSIZE (usually 512) are not 2968 * supported due to the page granularity bits (m->valid, 2969 * m->dirty, etc...). 2970 * 2971 * See man buf(9) for more information 2972 */ 2973 obj = bp->b_bufobj->bo_object; 2974 resid = bp->b_bufsize; 2975 poffset = bp->b_offset & PAGE_MASK; 2976 VM_OBJECT_WLOCK(obj); 2977 for (i = 0; i < bp->b_npages; i++) { 2978 m = bp->b_pages[i]; 2979 if (m == bogus_page) 2980 panic("vfs_vmio_invalidate: Unexpected bogus page."); 2981 bp->b_pages[i] = NULL; 2982 2983 presid = resid > (PAGE_SIZE - poffset) ? 2984 (PAGE_SIZE - poffset) : resid; 2985 KASSERT(presid >= 0, ("brelse: extra page")); 2986 while (vm_page_xbusied(m)) { 2987 vm_page_lock(m); 2988 VM_OBJECT_WUNLOCK(obj); 2989 vm_page_busy_sleep(m, "mbncsh", true); 2990 VM_OBJECT_WLOCK(obj); 2991 } 2992 if (pmap_page_wired_mappings(m) == 0) 2993 vm_page_set_invalid(m, poffset, presid); 2994 vfs_vmio_unwire(bp, m); 2995 resid -= presid; 2996 poffset = 0; 2997 } 2998 VM_OBJECT_WUNLOCK(obj); 2999 bp->b_npages = 0; 3000 } 3001 3002 /* 3003 * Page-granular truncation of an existing VMIO buffer. 3004 */ 3005 static void 3006 vfs_vmio_truncate(struct buf *bp, int desiredpages) 3007 { 3008 vm_object_t obj; 3009 vm_page_t m; 3010 int i; 3011 3012 if (bp->b_npages == desiredpages) 3013 return; 3014 3015 if (buf_mapped(bp)) { 3016 BUF_CHECK_MAPPED(bp); 3017 pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + 3018 (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); 3019 } else 3020 BUF_CHECK_UNMAPPED(bp); 3021 3022 /* 3023 * The object lock is needed only if we will attempt to free pages. 3024 */ 3025 obj = (bp->b_flags & B_DIRECT) != 0 ? bp->b_bufobj->bo_object : NULL; 3026 if (obj != NULL) 3027 VM_OBJECT_WLOCK(obj); 3028 for (i = desiredpages; i < bp->b_npages; i++) { 3029 m = bp->b_pages[i]; 3030 KASSERT(m != bogus_page, ("allocbuf: bogus page found")); 3031 bp->b_pages[i] = NULL; 3032 vfs_vmio_unwire(bp, m); 3033 } 3034 if (obj != NULL) 3035 VM_OBJECT_WUNLOCK(obj); 3036 bp->b_npages = desiredpages; 3037 } 3038 3039 /* 3040 * Byte granular extension of VMIO buffers. 3041 */ 3042 static void 3043 vfs_vmio_extend(struct buf *bp, int desiredpages, int size) 3044 { 3045 /* 3046 * We are growing the buffer, possibly in a 3047 * byte-granular fashion. 3048 */ 3049 vm_object_t obj; 3050 vm_offset_t toff; 3051 vm_offset_t tinc; 3052 vm_page_t m; 3053 3054 /* 3055 * Step 1, bring in the VM pages from the object, allocating 3056 * them if necessary. We must clear B_CACHE if these pages 3057 * are not valid for the range covered by the buffer. 3058 */ 3059 obj = bp->b_bufobj->bo_object; 3060 VM_OBJECT_WLOCK(obj); 3061 if (bp->b_npages < desiredpages) { 3062 /* 3063 * We must allocate system pages since blocking 3064 * here could interfere with paging I/O, no 3065 * matter which process we are. 3066 * 3067 * Only exclusive busy can be tested here. 3068 * Blocking on shared busy might lead to 3069 * deadlocks once allocbuf() is called after 3070 * pages are vfs_busy_pages(). 3071 */ 3072 (void)vm_page_grab_pages(obj, 3073 OFF_TO_IDX(bp->b_offset) + bp->b_npages, 3074 VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | 3075 VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, 3076 &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); 3077 bp->b_npages = desiredpages; 3078 } 3079 3080 /* 3081 * Step 2. We've loaded the pages into the buffer, 3082 * we have to figure out if we can still have B_CACHE 3083 * set. Note that B_CACHE is set according to the 3084 * byte-granular range ( bcount and size ), not the 3085 * aligned range ( newbsize ). 3086 * 3087 * The VM test is against m->valid, which is DEV_BSIZE 3088 * aligned. Needless to say, the validity of the data 3089 * needs to also be DEV_BSIZE aligned. Note that this 3090 * fails with NFS if the server or some other client 3091 * extends the file's EOF. If our buffer is resized, 3092 * B_CACHE may remain set! XXX 3093 */ 3094 toff = bp->b_bcount; 3095 tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); 3096 while ((bp->b_flags & B_CACHE) && toff < size) { 3097 vm_pindex_t pi; 3098 3099 if (tinc > (size - toff)) 3100 tinc = size - toff; 3101 pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; 3102 m = bp->b_pages[pi]; 3103 vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); 3104 toff += tinc; 3105 tinc = PAGE_SIZE; 3106 } 3107 VM_OBJECT_WUNLOCK(obj); 3108 3109 /* 3110 * Step 3, fixup the KVA pmap. 3111 */ 3112 if (buf_mapped(bp)) 3113 bpmap_qenter(bp); 3114 else 3115 BUF_CHECK_UNMAPPED(bp); 3116 } 3117 3118 /* 3119 * Check to see if a block at a particular lbn is available for a clustered 3120 * write. 3121 */ 3122 static int 3123 vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) 3124 { 3125 struct buf *bpa; 3126 int match; 3127 3128 match = 0; 3129 3130 /* If the buf isn't in core skip it */ 3131 if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) 3132 return (0); 3133 3134 /* If the buf is busy we don't want to wait for it */ 3135 if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) 3136 return (0); 3137 3138 /* Only cluster with valid clusterable delayed write buffers */ 3139 if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != 3140 (B_DELWRI | B_CLUSTEROK)) 3141 goto done; 3142 3143 if (bpa->b_bufsize != size) 3144 goto done; 3145 3146 /* 3147 * Check to see if it is in the expected place on disk and that the 3148 * block has been mapped. 3149 */ 3150 if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) 3151 match = 1; 3152 done: 3153 BUF_UNLOCK(bpa); 3154 return (match); 3155 } 3156 3157 /* 3158 * vfs_bio_awrite: 3159 * 3160 * Implement clustered async writes for clearing out B_DELWRI buffers. 3161 * This is much better then the old way of writing only one buffer at 3162 * a time. Note that we may not be presented with the buffers in the 3163 * correct order, so we search for the cluster in both directions. 3164 */ 3165 int 3166 vfs_bio_awrite(struct buf *bp) 3167 { 3168 struct bufobj *bo; 3169 int i; 3170 int j; 3171 daddr_t lblkno = bp->b_lblkno; 3172 struct vnode *vp = bp->b_vp; 3173 int ncl; 3174 int nwritten; 3175 int size; 3176 int maxcl; 3177 int gbflags; 3178 3179 bo = &vp->v_bufobj; 3180 gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; 3181 /* 3182 * right now we support clustered writing only to regular files. If 3183 * we find a clusterable block we could be in the middle of a cluster 3184 * rather then at the beginning. 3185 */ 3186 if ((vp->v_type == VREG) && 3187 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 3188 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 3189 3190 size = vp->v_mount->mnt_stat.f_iosize; 3191 maxcl = MAXPHYS / size; 3192 3193 BO_RLOCK(bo); 3194 for (i = 1; i < maxcl; i++) 3195 if (vfs_bio_clcheck(vp, size, lblkno + i, 3196 bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) 3197 break; 3198 3199 for (j = 1; i + j <= maxcl && j <= lblkno; j++) 3200 if (vfs_bio_clcheck(vp, size, lblkno - j, 3201 bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) 3202 break; 3203 BO_RUNLOCK(bo); 3204 --j; 3205 ncl = i + j; 3206 /* 3207 * this is a possible cluster write 3208 */ 3209 if (ncl != 1) { 3210 BUF_UNLOCK(bp); 3211 nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, 3212 gbflags); 3213 return (nwritten); 3214 } 3215 } 3216 bremfree(bp); 3217 bp->b_flags |= B_ASYNC; 3218 /* 3219 * default (old) behavior, writing out only one block 3220 * 3221 * XXX returns b_bufsize instead of b_bcount for nwritten? 3222 */ 3223 nwritten = bp->b_bufsize; 3224 (void) bwrite(bp); 3225 3226 return (nwritten); 3227 } 3228 3229 /* 3230 * getnewbuf_kva: 3231 * 3232 * Allocate KVA for an empty buf header according to gbflags. 3233 */ 3234 static int 3235 getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) 3236 { 3237 3238 if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { 3239 /* 3240 * In order to keep fragmentation sane we only allocate kva 3241 * in BKVASIZE chunks. XXX with vmem we can do page size. 3242 */ 3243 maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; 3244 3245 if (maxsize != bp->b_kvasize && 3246 bufkva_alloc(bp, maxsize, gbflags)) 3247 return (ENOSPC); 3248 } 3249 return (0); 3250 } 3251 3252 /* 3253 * getnewbuf: 3254 * 3255 * Find and initialize a new buffer header, freeing up existing buffers 3256 * in the bufqueues as necessary. The new buffer is returned locked. 3257 * 3258 * We block if: 3259 * We have insufficient buffer headers 3260 * We have insufficient buffer space 3261 * buffer_arena is too fragmented ( space reservation fails ) 3262 * If we have to flush dirty buffers ( but we try to avoid this ) 3263 * 3264 * The caller is responsible for releasing the reserved bufspace after 3265 * allocbuf() is called. 3266 */ 3267 static struct buf * 3268 getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) 3269 { 3270 struct bufdomain *bd; 3271 struct buf *bp; 3272 bool metadata, reserved; 3273 3274 bp = NULL; 3275 KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3276 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3277 if (!unmapped_buf_allowed) 3278 gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3279 3280 if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || 3281 vp->v_type == VCHR) 3282 metadata = true; 3283 else 3284 metadata = false; 3285 if (vp == NULL) 3286 bd = &bdomain[0]; 3287 else 3288 bd = &bdomain[vp->v_bufobj.bo_domain]; 3289 3290 counter_u64_add(getnewbufcalls, 1); 3291 reserved = false; 3292 do { 3293 if (reserved == false && 3294 bufspace_reserve(bd, maxsize, metadata) != 0) { 3295 counter_u64_add(getnewbufrestarts, 1); 3296 continue; 3297 } 3298 reserved = true; 3299 if ((bp = buf_alloc(bd)) == NULL) { 3300 counter_u64_add(getnewbufrestarts, 1); 3301 continue; 3302 } 3303 if (getnewbuf_kva(bp, gbflags, maxsize) == 0) 3304 return (bp); 3305 break; 3306 } while (buf_recycle(bd, false) == 0); 3307 3308 if (reserved) 3309 bufspace_release(bd, maxsize); 3310 if (bp != NULL) { 3311 bp->b_flags |= B_INVAL; 3312 brelse(bp); 3313 } 3314 bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); 3315 3316 return (NULL); 3317 } 3318 3319 /* 3320 * buf_daemon: 3321 * 3322 * buffer flushing daemon. Buffers are normally flushed by the 3323 * update daemon but if it cannot keep up this process starts to 3324 * take the load in an attempt to prevent getnewbuf() from blocking. 3325 */ 3326 static struct kproc_desc buf_kp = { 3327 "bufdaemon", 3328 buf_daemon, 3329 &bufdaemonproc 3330 }; 3331 SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); 3332 3333 static int 3334 buf_flush(struct vnode *vp, struct bufdomain *bd, int target) 3335 { 3336 int flushed; 3337 3338 flushed = flushbufqueues(vp, bd, target, 0); 3339 if (flushed == 0) { 3340 /* 3341 * Could not find any buffers without rollback 3342 * dependencies, so just write the first one 3343 * in the hopes of eventually making progress. 3344 */ 3345 if (vp != NULL && target > 2) 3346 target /= 2; 3347 flushbufqueues(vp, bd, target, 1); 3348 } 3349 return (flushed); 3350 } 3351 3352 static void 3353 buf_daemon() 3354 { 3355 struct bufdomain *bd; 3356 int speedupreq; 3357 int lodirty; 3358 int i; 3359 3360 /* 3361 * This process needs to be suspended prior to shutdown sync. 3362 */ 3363 EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, 3364 SHUTDOWN_PRI_LAST + 100); 3365 3366 /* 3367 * Start the buf clean daemons as children threads. 3368 */ 3369 for (i = 0 ; i < buf_domains; i++) { 3370 int error; 3371 3372 error = kthread_add((void (*)(void *))bufspace_daemon, 3373 &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); 3374 if (error) 3375 panic("error %d spawning bufspace daemon", error); 3376 } 3377 3378 /* 3379 * This process is allowed to take the buffer cache to the limit 3380 */ 3381 curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; 3382 mtx_lock(&bdlock); 3383 for (;;) { 3384 bd_request = 0; 3385 mtx_unlock(&bdlock); 3386 3387 kthread_suspend_check(); 3388 3389 /* 3390 * Save speedupreq for this pass and reset to capture new 3391 * requests. 3392 */ 3393 speedupreq = bd_speedupreq; 3394 bd_speedupreq = 0; 3395 3396 /* 3397 * Flush each domain sequentially according to its level and 3398 * the speedup request. 3399 */ 3400 for (i = 0; i < buf_domains; i++) { 3401 bd = &bdomain[i]; 3402 if (speedupreq) 3403 lodirty = bd->bd_numdirtybuffers / 2; 3404 else 3405 lodirty = bd->bd_lodirtybuffers; 3406 while (bd->bd_numdirtybuffers > lodirty) { 3407 if (buf_flush(NULL, bd, 3408 bd->bd_numdirtybuffers - lodirty) == 0) 3409 break; 3410 kern_yield(PRI_USER); 3411 } 3412 } 3413 3414 /* 3415 * Only clear bd_request if we have reached our low water 3416 * mark. The buf_daemon normally waits 1 second and 3417 * then incrementally flushes any dirty buffers that have 3418 * built up, within reason. 3419 * 3420 * If we were unable to hit our low water mark and couldn't 3421 * find any flushable buffers, we sleep for a short period 3422 * to avoid endless loops on unlockable buffers. 3423 */ 3424 mtx_lock(&bdlock); 3425 if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { 3426 /* 3427 * We reached our low water mark, reset the 3428 * request and sleep until we are needed again. 3429 * The sleep is just so the suspend code works. 3430 */ 3431 bd_request = 0; 3432 /* 3433 * Do an extra wakeup in case dirty threshold 3434 * changed via sysctl and the explicit transition 3435 * out of shortfall was missed. 3436 */ 3437 bdirtywakeup(); 3438 if (runningbufspace <= lorunningspace) 3439 runningwakeup(); 3440 msleep(&bd_request, &bdlock, PVM, "psleep", hz); 3441 } else { 3442 /* 3443 * We couldn't find any flushable dirty buffers but 3444 * still have too many dirty buffers, we 3445 * have to sleep and try again. (rare) 3446 */ 3447 msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); 3448 } 3449 } 3450 } 3451 3452 /* 3453 * flushbufqueues: 3454 * 3455 * Try to flush a buffer in the dirty queue. We must be careful to 3456 * free up B_INVAL buffers instead of write them, which NFS is 3457 * particularly sensitive to. 3458 */ 3459 static int flushwithdeps = 0; 3460 SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW, &flushwithdeps, 3461 0, "Number of buffers flushed with dependecies that require rollbacks"); 3462 3463 static int 3464 flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, 3465 int flushdeps) 3466 { 3467 struct bufqueue *bq; 3468 struct buf *sentinel; 3469 struct vnode *vp; 3470 struct mount *mp; 3471 struct buf *bp; 3472 int hasdeps; 3473 int flushed; 3474 int error; 3475 bool unlock; 3476 3477 flushed = 0; 3478 bq = &bd->bd_dirtyq; 3479 bp = NULL; 3480 sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); 3481 sentinel->b_qindex = QUEUE_SENTINEL; 3482 BQ_LOCK(bq); 3483 TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); 3484 BQ_UNLOCK(bq); 3485 while (flushed != target) { 3486 maybe_yield(); 3487 BQ_LOCK(bq); 3488 bp = TAILQ_NEXT(sentinel, b_freelist); 3489 if (bp != NULL) { 3490 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3491 TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, 3492 b_freelist); 3493 } else { 3494 BQ_UNLOCK(bq); 3495 break; 3496 } 3497 /* 3498 * Skip sentinels inserted by other invocations of the 3499 * flushbufqueues(), taking care to not reorder them. 3500 * 3501 * Only flush the buffers that belong to the 3502 * vnode locked by the curthread. 3503 */ 3504 if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && 3505 bp->b_vp != lvp)) { 3506 BQ_UNLOCK(bq); 3507 continue; 3508 } 3509 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); 3510 BQ_UNLOCK(bq); 3511 if (error != 0) 3512 continue; 3513 3514 /* 3515 * BKGRDINPROG can only be set with the buf and bufobj 3516 * locks both held. We tolerate a race to clear it here. 3517 */ 3518 if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || 3519 (bp->b_flags & B_DELWRI) == 0) { 3520 BUF_UNLOCK(bp); 3521 continue; 3522 } 3523 if (bp->b_flags & B_INVAL) { 3524 bremfreef(bp); 3525 brelse(bp); 3526 flushed++; 3527 continue; 3528 } 3529 3530 if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { 3531 if (flushdeps == 0) { 3532 BUF_UNLOCK(bp); 3533 continue; 3534 } 3535 hasdeps = 1; 3536 } else 3537 hasdeps = 0; 3538 /* 3539 * We must hold the lock on a vnode before writing 3540 * one of its buffers. Otherwise we may confuse, or 3541 * in the case of a snapshot vnode, deadlock the 3542 * system. 3543 * 3544 * The lock order here is the reverse of the normal 3545 * of vnode followed by buf lock. This is ok because 3546 * the NOWAIT will prevent deadlock. 3547 */ 3548 vp = bp->b_vp; 3549 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { 3550 BUF_UNLOCK(bp); 3551 continue; 3552 } 3553 if (lvp == NULL) { 3554 unlock = true; 3555 error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); 3556 } else { 3557 ASSERT_VOP_LOCKED(vp, "getbuf"); 3558 unlock = false; 3559 error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : 3560 vn_lock(vp, LK_TRYUPGRADE); 3561 } 3562 if (error == 0) { 3563 CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", 3564 bp, bp->b_vp, bp->b_flags); 3565 if (curproc == bufdaemonproc) { 3566 vfs_bio_awrite(bp); 3567 } else { 3568 bremfree(bp); 3569 bwrite(bp); 3570 counter_u64_add(notbufdflushes, 1); 3571 } 3572 vn_finished_write(mp); 3573 if (unlock) 3574 VOP_UNLOCK(vp, 0); 3575 flushwithdeps += hasdeps; 3576 flushed++; 3577 3578 /* 3579 * Sleeping on runningbufspace while holding 3580 * vnode lock leads to deadlock. 3581 */ 3582 if (curproc == bufdaemonproc && 3583 runningbufspace > hirunningspace) 3584 waitrunningbufspace(); 3585 continue; 3586 } 3587 vn_finished_write(mp); 3588 BUF_UNLOCK(bp); 3589 } 3590 BQ_LOCK(bq); 3591 TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); 3592 BQ_UNLOCK(bq); 3593 free(sentinel, M_TEMP); 3594 return (flushed); 3595 } 3596 3597 /* 3598 * Check to see if a block is currently memory resident. 3599 */ 3600 struct buf * 3601 incore(struct bufobj *bo, daddr_t blkno) 3602 { 3603 struct buf *bp; 3604 3605 BO_RLOCK(bo); 3606 bp = gbincore(bo, blkno); 3607 BO_RUNLOCK(bo); 3608 return (bp); 3609 } 3610 3611 /* 3612 * Returns true if no I/O is needed to access the 3613 * associated VM object. This is like incore except 3614 * it also hunts around in the VM system for the data. 3615 */ 3616 3617 static int 3618 inmem(struct vnode * vp, daddr_t blkno) 3619 { 3620 vm_object_t obj; 3621 vm_offset_t toff, tinc, size; 3622 vm_page_t m; 3623 vm_ooffset_t off; 3624 3625 ASSERT_VOP_LOCKED(vp, "inmem"); 3626 3627 if (incore(&vp->v_bufobj, blkno)) 3628 return 1; 3629 if (vp->v_mount == NULL) 3630 return 0; 3631 obj = vp->v_object; 3632 if (obj == NULL) 3633 return (0); 3634 3635 size = PAGE_SIZE; 3636 if (size > vp->v_mount->mnt_stat.f_iosize) 3637 size = vp->v_mount->mnt_stat.f_iosize; 3638 off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; 3639 3640 VM_OBJECT_RLOCK(obj); 3641 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 3642 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 3643 if (!m) 3644 goto notinmem; 3645 tinc = size; 3646 if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) 3647 tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); 3648 if (vm_page_is_valid(m, 3649 (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) 3650 goto notinmem; 3651 } 3652 VM_OBJECT_RUNLOCK(obj); 3653 return 1; 3654 3655 notinmem: 3656 VM_OBJECT_RUNLOCK(obj); 3657 return (0); 3658 } 3659 3660 /* 3661 * Set the dirty range for a buffer based on the status of the dirty 3662 * bits in the pages comprising the buffer. The range is limited 3663 * to the size of the buffer. 3664 * 3665 * Tell the VM system that the pages associated with this buffer 3666 * are clean. This is used for delayed writes where the data is 3667 * going to go to disk eventually without additional VM intevention. 3668 * 3669 * Note that while we only really need to clean through to b_bcount, we 3670 * just go ahead and clean through to b_bufsize. 3671 */ 3672 static void 3673 vfs_clean_pages_dirty_buf(struct buf *bp) 3674 { 3675 vm_ooffset_t foff, noff, eoff; 3676 vm_page_t m; 3677 int i; 3678 3679 if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) 3680 return; 3681 3682 foff = bp->b_offset; 3683 KASSERT(bp->b_offset != NOOFFSET, 3684 ("vfs_clean_pages_dirty_buf: no buffer offset")); 3685 3686 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 3687 vfs_drain_busy_pages(bp); 3688 vfs_setdirty_locked_object(bp); 3689 for (i = 0; i < bp->b_npages; i++) { 3690 noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 3691 eoff = noff; 3692 if (eoff > bp->b_offset + bp->b_bufsize) 3693 eoff = bp->b_offset + bp->b_bufsize; 3694 m = bp->b_pages[i]; 3695 vfs_page_set_validclean(bp, foff, m); 3696 /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ 3697 foff = noff; 3698 } 3699 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 3700 } 3701 3702 static void 3703 vfs_setdirty_locked_object(struct buf *bp) 3704 { 3705 vm_object_t object; 3706 int i; 3707 3708 object = bp->b_bufobj->bo_object; 3709 VM_OBJECT_ASSERT_WLOCKED(object); 3710 3711 /* 3712 * We qualify the scan for modified pages on whether the 3713 * object has been flushed yet. 3714 */ 3715 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0) { 3716 vm_offset_t boffset; 3717 vm_offset_t eoffset; 3718 3719 /* 3720 * test the pages to see if they have been modified directly 3721 * by users through the VM system. 3722 */ 3723 for (i = 0; i < bp->b_npages; i++) 3724 vm_page_test_dirty(bp->b_pages[i]); 3725 3726 /* 3727 * Calculate the encompassing dirty range, boffset and eoffset, 3728 * (eoffset - boffset) bytes. 3729 */ 3730 3731 for (i = 0; i < bp->b_npages; i++) { 3732 if (bp->b_pages[i]->dirty) 3733 break; 3734 } 3735 boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3736 3737 for (i = bp->b_npages - 1; i >= 0; --i) { 3738 if (bp->b_pages[i]->dirty) { 3739 break; 3740 } 3741 } 3742 eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); 3743 3744 /* 3745 * Fit it to the buffer. 3746 */ 3747 3748 if (eoffset > bp->b_bcount) 3749 eoffset = bp->b_bcount; 3750 3751 /* 3752 * If we have a good dirty range, merge with the existing 3753 * dirty range. 3754 */ 3755 3756 if (boffset < eoffset) { 3757 if (bp->b_dirtyoff > boffset) 3758 bp->b_dirtyoff = boffset; 3759 if (bp->b_dirtyend < eoffset) 3760 bp->b_dirtyend = eoffset; 3761 } 3762 } 3763 } 3764 3765 /* 3766 * Allocate the KVA mapping for an existing buffer. 3767 * If an unmapped buffer is provided but a mapped buffer is requested, take 3768 * also care to properly setup mappings between pages and KVA. 3769 */ 3770 static void 3771 bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) 3772 { 3773 int bsize, maxsize, need_mapping, need_kva; 3774 off_t offset; 3775 3776 need_mapping = bp->b_data == unmapped_buf && 3777 (gbflags & GB_UNMAPPED) == 0; 3778 need_kva = bp->b_kvabase == unmapped_buf && 3779 bp->b_data == unmapped_buf && 3780 (gbflags & GB_KVAALLOC) != 0; 3781 if (!need_mapping && !need_kva) 3782 return; 3783 3784 BUF_CHECK_UNMAPPED(bp); 3785 3786 if (need_mapping && bp->b_kvabase != unmapped_buf) { 3787 /* 3788 * Buffer is not mapped, but the KVA was already 3789 * reserved at the time of the instantiation. Use the 3790 * allocated space. 3791 */ 3792 goto has_addr; 3793 } 3794 3795 /* 3796 * Calculate the amount of the address space we would reserve 3797 * if the buffer was mapped. 3798 */ 3799 bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; 3800 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 3801 offset = blkno * bsize; 3802 maxsize = size + (offset & PAGE_MASK); 3803 maxsize = imax(maxsize, bsize); 3804 3805 while (bufkva_alloc(bp, maxsize, gbflags) != 0) { 3806 if ((gbflags & GB_NOWAIT_BD) != 0) { 3807 /* 3808 * XXXKIB: defragmentation cannot 3809 * succeed, not sure what else to do. 3810 */ 3811 panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); 3812 } 3813 counter_u64_add(mappingrestarts, 1); 3814 bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); 3815 } 3816 has_addr: 3817 if (need_mapping) { 3818 /* b_offset is handled by bpmap_qenter. */ 3819 bp->b_data = bp->b_kvabase; 3820 BUF_CHECK_MAPPED(bp); 3821 bpmap_qenter(bp); 3822 } 3823 } 3824 3825 /* 3826 * getblk: 3827 * 3828 * Get a block given a specified block and offset into a file/device. 3829 * The buffers B_DONE bit will be cleared on return, making it almost 3830 * ready for an I/O initiation. B_INVAL may or may not be set on 3831 * return. The caller should clear B_INVAL prior to initiating a 3832 * READ. 3833 * 3834 * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for 3835 * an existing buffer. 3836 * 3837 * For a VMIO buffer, B_CACHE is modified according to the backing VM. 3838 * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set 3839 * and then cleared based on the backing VM. If the previous buffer is 3840 * non-0-sized but invalid, B_CACHE will be cleared. 3841 * 3842 * If getblk() must create a new buffer, the new buffer is returned with 3843 * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which 3844 * case it is returned with B_INVAL clear and B_CACHE set based on the 3845 * backing VM. 3846 * 3847 * getblk() also forces a bwrite() for any B_DELWRI buffer whos 3848 * B_CACHE bit is clear. 3849 * 3850 * What this means, basically, is that the caller should use B_CACHE to 3851 * determine whether the buffer is fully valid or not and should clear 3852 * B_INVAL prior to issuing a read. If the caller intends to validate 3853 * the buffer by loading its data area with something, the caller needs 3854 * to clear B_INVAL. If the caller does this without issuing an I/O, 3855 * the caller should set B_CACHE ( as an optimization ), else the caller 3856 * should issue the I/O and biodone() will set B_CACHE if the I/O was 3857 * a write attempt or if it was a successful read. If the caller 3858 * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR 3859 * prior to issuing the READ. biodone() will *not* clear B_INVAL. 3860 */ 3861 struct buf * 3862 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, 3863 int flags) 3864 { 3865 struct buf *bp; 3866 struct bufobj *bo; 3867 int bsize, error, maxsize, vmio; 3868 off_t offset; 3869 3870 CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); 3871 KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, 3872 ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); 3873 ASSERT_VOP_LOCKED(vp, "getblk"); 3874 if (size > maxbcachebuf) 3875 panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, 3876 maxbcachebuf); 3877 if (!unmapped_buf_allowed) 3878 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 3879 3880 bo = &vp->v_bufobj; 3881 loop: 3882 BO_RLOCK(bo); 3883 bp = gbincore(bo, blkno); 3884 if (bp != NULL) { 3885 int lockflags; 3886 /* 3887 * Buffer is in-core. If the buffer is not busy nor managed, 3888 * it must be on a queue. 3889 */ 3890 lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; 3891 3892 if (flags & GB_LOCK_NOWAIT) 3893 lockflags |= LK_NOWAIT; 3894 3895 error = BUF_TIMELOCK(bp, lockflags, 3896 BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); 3897 3898 /* 3899 * If we slept and got the lock we have to restart in case 3900 * the buffer changed identities. 3901 */ 3902 if (error == ENOLCK) 3903 goto loop; 3904 /* We timed out or were interrupted. */ 3905 else if (error) 3906 return (NULL); 3907 /* If recursed, assume caller knows the rules. */ 3908 else if (BUF_LOCKRECURSED(bp)) 3909 goto end; 3910 3911 /* 3912 * The buffer is locked. B_CACHE is cleared if the buffer is 3913 * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set 3914 * and for a VMIO buffer B_CACHE is adjusted according to the 3915 * backing VM cache. 3916 */ 3917 if (bp->b_flags & B_INVAL) 3918 bp->b_flags &= ~B_CACHE; 3919 else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) 3920 bp->b_flags |= B_CACHE; 3921 if (bp->b_flags & B_MANAGED) 3922 MPASS(bp->b_qindex == QUEUE_NONE); 3923 else 3924 bremfree(bp); 3925 3926 /* 3927 * check for size inconsistencies for non-VMIO case. 3928 */ 3929 if (bp->b_bcount != size) { 3930 if ((bp->b_flags & B_VMIO) == 0 || 3931 (size > bp->b_kvasize)) { 3932 if (bp->b_flags & B_DELWRI) { 3933 bp->b_flags |= B_NOCACHE; 3934 bwrite(bp); 3935 } else { 3936 if (LIST_EMPTY(&bp->b_dep)) { 3937 bp->b_flags |= B_RELBUF; 3938 brelse(bp); 3939 } else { 3940 bp->b_flags |= B_NOCACHE; 3941 bwrite(bp); 3942 } 3943 } 3944 goto loop; 3945 } 3946 } 3947 3948 /* 3949 * Handle the case of unmapped buffer which should 3950 * become mapped, or the buffer for which KVA 3951 * reservation is requested. 3952 */ 3953 bp_unmapped_get_kva(bp, blkno, size, flags); 3954 3955 /* 3956 * If the size is inconsistent in the VMIO case, we can resize 3957 * the buffer. This might lead to B_CACHE getting set or 3958 * cleared. If the size has not changed, B_CACHE remains 3959 * unchanged from its previous state. 3960 */ 3961 allocbuf(bp, size); 3962 3963 KASSERT(bp->b_offset != NOOFFSET, 3964 ("getblk: no buffer offset")); 3965 3966 /* 3967 * A buffer with B_DELWRI set and B_CACHE clear must 3968 * be committed before we can return the buffer in 3969 * order to prevent the caller from issuing a read 3970 * ( due to B_CACHE not being set ) and overwriting 3971 * it. 3972 * 3973 * Most callers, including NFS and FFS, need this to 3974 * operate properly either because they assume they 3975 * can issue a read if B_CACHE is not set, or because 3976 * ( for example ) an uncached B_DELWRI might loop due 3977 * to softupdates re-dirtying the buffer. In the latter 3978 * case, B_CACHE is set after the first write completes, 3979 * preventing further loops. 3980 * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE 3981 * above while extending the buffer, we cannot allow the 3982 * buffer to remain with B_CACHE set after the write 3983 * completes or it will represent a corrupt state. To 3984 * deal with this we set B_NOCACHE to scrap the buffer 3985 * after the write. 3986 * 3987 * We might be able to do something fancy, like setting 3988 * B_CACHE in bwrite() except if B_DELWRI is already set, 3989 * so the below call doesn't set B_CACHE, but that gets real 3990 * confusing. This is much easier. 3991 */ 3992 3993 if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { 3994 bp->b_flags |= B_NOCACHE; 3995 bwrite(bp); 3996 goto loop; 3997 } 3998 bp->b_flags &= ~B_DONE; 3999 } else { 4000 /* 4001 * Buffer is not in-core, create new buffer. The buffer 4002 * returned by getnewbuf() is locked. Note that the returned 4003 * buffer is also considered valid (not marked B_INVAL). 4004 */ 4005 BO_RUNLOCK(bo); 4006 /* 4007 * If the user does not want us to create the buffer, bail out 4008 * here. 4009 */ 4010 if (flags & GB_NOCREAT) 4011 return NULL; 4012 if (bdomain[bo->bo_domain].bd_freebuffers == 0 && 4013 TD_IS_IDLETHREAD(curthread)) 4014 return NULL; 4015 4016 bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; 4017 KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); 4018 offset = blkno * bsize; 4019 vmio = vp->v_object != NULL; 4020 if (vmio) { 4021 maxsize = size + (offset & PAGE_MASK); 4022 } else { 4023 maxsize = size; 4024 /* Do not allow non-VMIO notmapped buffers. */ 4025 flags &= ~(GB_UNMAPPED | GB_KVAALLOC); 4026 } 4027 maxsize = imax(maxsize, bsize); 4028 4029 bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); 4030 if (bp == NULL) { 4031 if (slpflag || slptimeo) 4032 return NULL; 4033 /* 4034 * XXX This is here until the sleep path is diagnosed 4035 * enough to work under very low memory conditions. 4036 * 4037 * There's an issue on low memory, 4BSD+non-preempt 4038 * systems (eg MIPS routers with 32MB RAM) where buffer 4039 * exhaustion occurs without sleeping for buffer 4040 * reclaimation. This just sticks in a loop and 4041 * constantly attempts to allocate a buffer, which 4042 * hits exhaustion and tries to wakeup bufdaemon. 4043 * This never happens because we never yield. 4044 * 4045 * The real solution is to identify and fix these cases 4046 * so we aren't effectively busy-waiting in a loop 4047 * until the reclaimation path has cycles to run. 4048 */ 4049 kern_yield(PRI_USER); 4050 goto loop; 4051 } 4052 4053 /* 4054 * This code is used to make sure that a buffer is not 4055 * created while the getnewbuf routine is blocked. 4056 * This can be a problem whether the vnode is locked or not. 4057 * If the buffer is created out from under us, we have to 4058 * throw away the one we just created. 4059 * 4060 * Note: this must occur before we associate the buffer 4061 * with the vp especially considering limitations in 4062 * the splay tree implementation when dealing with duplicate 4063 * lblkno's. 4064 */ 4065 BO_LOCK(bo); 4066 if (gbincore(bo, blkno)) { 4067 BO_UNLOCK(bo); 4068 bp->b_flags |= B_INVAL; 4069 bufspace_release(bufdomain(bp), maxsize); 4070 brelse(bp); 4071 goto loop; 4072 } 4073 4074 /* 4075 * Insert the buffer into the hash, so that it can 4076 * be found by incore. 4077 */ 4078 bp->b_blkno = bp->b_lblkno = blkno; 4079 bp->b_offset = offset; 4080 bgetvp(vp, bp); 4081 BO_UNLOCK(bo); 4082 4083 /* 4084 * set B_VMIO bit. allocbuf() the buffer bigger. Since the 4085 * buffer size starts out as 0, B_CACHE will be set by 4086 * allocbuf() for the VMIO case prior to it testing the 4087 * backing store for validity. 4088 */ 4089 4090 if (vmio) { 4091 bp->b_flags |= B_VMIO; 4092 KASSERT(vp->v_object == bp->b_bufobj->bo_object, 4093 ("ARGH! different b_bufobj->bo_object %p %p %p\n", 4094 bp, vp->v_object, bp->b_bufobj->bo_object)); 4095 } else { 4096 bp->b_flags &= ~B_VMIO; 4097 KASSERT(bp->b_bufobj->bo_object == NULL, 4098 ("ARGH! has b_bufobj->bo_object %p %p\n", 4099 bp, bp->b_bufobj->bo_object)); 4100 BUF_CHECK_MAPPED(bp); 4101 } 4102 4103 allocbuf(bp, size); 4104 bufspace_release(bufdomain(bp), maxsize); 4105 bp->b_flags &= ~B_DONE; 4106 } 4107 CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); 4108 BUF_ASSERT_HELD(bp); 4109 end: 4110 buf_track(bp, __func__); 4111 KASSERT(bp->b_bufobj == bo, 4112 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); 4113 return (bp); 4114 } 4115 4116 /* 4117 * Get an empty, disassociated buffer of given size. The buffer is initially 4118 * set to B_INVAL. 4119 */ 4120 struct buf * 4121 geteblk(int size, int flags) 4122 { 4123 struct buf *bp; 4124 int maxsize; 4125 4126 maxsize = (size + BKVAMASK) & ~BKVAMASK; 4127 while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { 4128 if ((flags & GB_NOWAIT_BD) && 4129 (curthread->td_pflags & TDP_BUFNEED) != 0) 4130 return (NULL); 4131 } 4132 allocbuf(bp, size); 4133 bufspace_release(bufdomain(bp), maxsize); 4134 bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ 4135 BUF_ASSERT_HELD(bp); 4136 return (bp); 4137 } 4138 4139 /* 4140 * Truncate the backing store for a non-vmio buffer. 4141 */ 4142 static void 4143 vfs_nonvmio_truncate(struct buf *bp, int newbsize) 4144 { 4145 4146 if (bp->b_flags & B_MALLOC) { 4147 /* 4148 * malloced buffers are not shrunk 4149 */ 4150 if (newbsize == 0) { 4151 bufmallocadjust(bp, 0); 4152 free(bp->b_data, M_BIOBUF); 4153 bp->b_data = bp->b_kvabase; 4154 bp->b_flags &= ~B_MALLOC; 4155 } 4156 return; 4157 } 4158 vm_hold_free_pages(bp, newbsize); 4159 bufspace_adjust(bp, newbsize); 4160 } 4161 4162 /* 4163 * Extend the backing for a non-VMIO buffer. 4164 */ 4165 static void 4166 vfs_nonvmio_extend(struct buf *bp, int newbsize) 4167 { 4168 caddr_t origbuf; 4169 int origbufsize; 4170 4171 /* 4172 * We only use malloced memory on the first allocation. 4173 * and revert to page-allocated memory when the buffer 4174 * grows. 4175 * 4176 * There is a potential smp race here that could lead 4177 * to bufmallocspace slightly passing the max. It 4178 * is probably extremely rare and not worth worrying 4179 * over. 4180 */ 4181 if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && 4182 bufmallocspace < maxbufmallocspace) { 4183 bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); 4184 bp->b_flags |= B_MALLOC; 4185 bufmallocadjust(bp, newbsize); 4186 return; 4187 } 4188 4189 /* 4190 * If the buffer is growing on its other-than-first 4191 * allocation then we revert to the page-allocation 4192 * scheme. 4193 */ 4194 origbuf = NULL; 4195 origbufsize = 0; 4196 if (bp->b_flags & B_MALLOC) { 4197 origbuf = bp->b_data; 4198 origbufsize = bp->b_bufsize; 4199 bp->b_data = bp->b_kvabase; 4200 bufmallocadjust(bp, 0); 4201 bp->b_flags &= ~B_MALLOC; 4202 newbsize = round_page(newbsize); 4203 } 4204 vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, 4205 (vm_offset_t) bp->b_data + newbsize); 4206 if (origbuf != NULL) { 4207 bcopy(origbuf, bp->b_data, origbufsize); 4208 free(origbuf, M_BIOBUF); 4209 } 4210 bufspace_adjust(bp, newbsize); 4211 } 4212 4213 /* 4214 * This code constitutes the buffer memory from either anonymous system 4215 * memory (in the case of non-VMIO operations) or from an associated 4216 * VM object (in the case of VMIO operations). This code is able to 4217 * resize a buffer up or down. 4218 * 4219 * Note that this code is tricky, and has many complications to resolve 4220 * deadlock or inconsistent data situations. Tread lightly!!! 4221 * There are B_CACHE and B_DELWRI interactions that must be dealt with by 4222 * the caller. Calling this code willy nilly can result in the loss of data. 4223 * 4224 * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with 4225 * B_CACHE for the non-VMIO case. 4226 */ 4227 int 4228 allocbuf(struct buf *bp, int size) 4229 { 4230 int newbsize; 4231 4232 BUF_ASSERT_HELD(bp); 4233 4234 if (bp->b_bcount == size) 4235 return (1); 4236 4237 if (bp->b_kvasize != 0 && bp->b_kvasize < size) 4238 panic("allocbuf: buffer too small"); 4239 4240 newbsize = roundup2(size, DEV_BSIZE); 4241 if ((bp->b_flags & B_VMIO) == 0) { 4242 if ((bp->b_flags & B_MALLOC) == 0) 4243 newbsize = round_page(newbsize); 4244 /* 4245 * Just get anonymous memory from the kernel. Don't 4246 * mess with B_CACHE. 4247 */ 4248 if (newbsize < bp->b_bufsize) 4249 vfs_nonvmio_truncate(bp, newbsize); 4250 else if (newbsize > bp->b_bufsize) 4251 vfs_nonvmio_extend(bp, newbsize); 4252 } else { 4253 int desiredpages; 4254 4255 desiredpages = (size == 0) ? 0 : 4256 num_pages((bp->b_offset & PAGE_MASK) + newbsize); 4257 4258 if (bp->b_flags & B_MALLOC) 4259 panic("allocbuf: VMIO buffer can't be malloced"); 4260 /* 4261 * Set B_CACHE initially if buffer is 0 length or will become 4262 * 0-length. 4263 */ 4264 if (size == 0 || bp->b_bufsize == 0) 4265 bp->b_flags |= B_CACHE; 4266 4267 if (newbsize < bp->b_bufsize) 4268 vfs_vmio_truncate(bp, desiredpages); 4269 /* XXX This looks as if it should be newbsize > b_bufsize */ 4270 else if (size > bp->b_bcount) 4271 vfs_vmio_extend(bp, desiredpages, size); 4272 bufspace_adjust(bp, newbsize); 4273 } 4274 bp->b_bcount = size; /* requested buffer size. */ 4275 return (1); 4276 } 4277 4278 extern int inflight_transient_maps; 4279 4280 void 4281 biodone(struct bio *bp) 4282 { 4283 struct mtx *mtxp; 4284 void (*done)(struct bio *); 4285 vm_offset_t start, end; 4286 4287 biotrack(bp, __func__); 4288 if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { 4289 bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; 4290 bp->bio_flags |= BIO_UNMAPPED; 4291 start = trunc_page((vm_offset_t)bp->bio_data); 4292 end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); 4293 bp->bio_data = unmapped_buf; 4294 pmap_qremove(start, atop(end - start)); 4295 vmem_free(transient_arena, start, end - start); 4296 atomic_add_int(&inflight_transient_maps, -1); 4297 } 4298 done = bp->bio_done; 4299 if (done == NULL) { 4300 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4301 mtx_lock(mtxp); 4302 bp->bio_flags |= BIO_DONE; 4303 wakeup(bp); 4304 mtx_unlock(mtxp); 4305 } else 4306 done(bp); 4307 } 4308 4309 /* 4310 * Wait for a BIO to finish. 4311 */ 4312 int 4313 biowait(struct bio *bp, const char *wchan) 4314 { 4315 struct mtx *mtxp; 4316 4317 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4318 mtx_lock(mtxp); 4319 while ((bp->bio_flags & BIO_DONE) == 0) 4320 msleep(bp, mtxp, PRIBIO, wchan, 0); 4321 mtx_unlock(mtxp); 4322 if (bp->bio_error != 0) 4323 return (bp->bio_error); 4324 if (!(bp->bio_flags & BIO_ERROR)) 4325 return (0); 4326 return (EIO); 4327 } 4328 4329 void 4330 biofinish(struct bio *bp, struct devstat *stat, int error) 4331 { 4332 4333 if (error) { 4334 bp->bio_error = error; 4335 bp->bio_flags |= BIO_ERROR; 4336 } 4337 if (stat != NULL) 4338 devstat_end_transaction_bio(stat, bp); 4339 biodone(bp); 4340 } 4341 4342 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) 4343 void 4344 biotrack_buf(struct bio *bp, const char *location) 4345 { 4346 4347 buf_track(bp->bio_track_bp, location); 4348 } 4349 #endif 4350 4351 /* 4352 * bufwait: 4353 * 4354 * Wait for buffer I/O completion, returning error status. The buffer 4355 * is left locked and B_DONE on return. B_EINTR is converted into an EINTR 4356 * error and cleared. 4357 */ 4358 int 4359 bufwait(struct buf *bp) 4360 { 4361 if (bp->b_iocmd == BIO_READ) 4362 bwait(bp, PRIBIO, "biord"); 4363 else 4364 bwait(bp, PRIBIO, "biowr"); 4365 if (bp->b_flags & B_EINTR) { 4366 bp->b_flags &= ~B_EINTR; 4367 return (EINTR); 4368 } 4369 if (bp->b_ioflags & BIO_ERROR) { 4370 return (bp->b_error ? bp->b_error : EIO); 4371 } else { 4372 return (0); 4373 } 4374 } 4375 4376 /* 4377 * bufdone: 4378 * 4379 * Finish I/O on a buffer, optionally calling a completion function. 4380 * This is usually called from an interrupt so process blocking is 4381 * not allowed. 4382 * 4383 * biodone is also responsible for setting B_CACHE in a B_VMIO bp. 4384 * In a non-VMIO bp, B_CACHE will be set on the next getblk() 4385 * assuming B_INVAL is clear. 4386 * 4387 * For the VMIO case, we set B_CACHE if the op was a read and no 4388 * read error occurred, or if the op was a write. B_CACHE is never 4389 * set if the buffer is invalid or otherwise uncacheable. 4390 * 4391 * biodone does not mess with B_INVAL, allowing the I/O routine or the 4392 * initiator to leave B_INVAL set to brelse the buffer out of existence 4393 * in the biodone routine. 4394 */ 4395 void 4396 bufdone(struct buf *bp) 4397 { 4398 struct bufobj *dropobj; 4399 void (*biodone)(struct buf *); 4400 4401 buf_track(bp, __func__); 4402 CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); 4403 dropobj = NULL; 4404 4405 KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); 4406 BUF_ASSERT_HELD(bp); 4407 4408 runningbufwakeup(bp); 4409 if (bp->b_iocmd == BIO_WRITE) 4410 dropobj = bp->b_bufobj; 4411 /* call optional completion function if requested */ 4412 if (bp->b_iodone != NULL) { 4413 biodone = bp->b_iodone; 4414 bp->b_iodone = NULL; 4415 (*biodone) (bp); 4416 if (dropobj) 4417 bufobj_wdrop(dropobj); 4418 return; 4419 } 4420 if (bp->b_flags & B_VMIO) { 4421 /* 4422 * Set B_CACHE if the op was a normal read and no error 4423 * occurred. B_CACHE is set for writes in the b*write() 4424 * routines. 4425 */ 4426 if (bp->b_iocmd == BIO_READ && 4427 !(bp->b_flags & (B_INVAL|B_NOCACHE)) && 4428 !(bp->b_ioflags & BIO_ERROR)) 4429 bp->b_flags |= B_CACHE; 4430 vfs_vmio_iodone(bp); 4431 } 4432 if (!LIST_EMPTY(&bp->b_dep)) 4433 buf_complete(bp); 4434 if ((bp->b_flags & B_CKHASH) != 0) { 4435 KASSERT(bp->b_iocmd == BIO_READ, 4436 ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); 4437 KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); 4438 (*bp->b_ckhashcalc)(bp); 4439 } 4440 /* 4441 * For asynchronous completions, release the buffer now. The brelse 4442 * will do a wakeup there if necessary - so no need to do a wakeup 4443 * here in the async case. The sync case always needs to do a wakeup. 4444 */ 4445 if (bp->b_flags & B_ASYNC) { 4446 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || 4447 (bp->b_ioflags & BIO_ERROR)) 4448 brelse(bp); 4449 else 4450 bqrelse(bp); 4451 } else 4452 bdone(bp); 4453 if (dropobj) 4454 bufobj_wdrop(dropobj); 4455 } 4456 4457 /* 4458 * This routine is called in lieu of iodone in the case of 4459 * incomplete I/O. This keeps the busy status for pages 4460 * consistent. 4461 */ 4462 void 4463 vfs_unbusy_pages(struct buf *bp) 4464 { 4465 int i; 4466 vm_object_t obj; 4467 vm_page_t m; 4468 4469 runningbufwakeup(bp); 4470 if (!(bp->b_flags & B_VMIO)) 4471 return; 4472 4473 obj = bp->b_bufobj->bo_object; 4474 VM_OBJECT_WLOCK(obj); 4475 for (i = 0; i < bp->b_npages; i++) { 4476 m = bp->b_pages[i]; 4477 if (m == bogus_page) { 4478 m = vm_page_lookup(obj, OFF_TO_IDX(bp->b_offset) + i); 4479 if (!m) 4480 panic("vfs_unbusy_pages: page missing\n"); 4481 bp->b_pages[i] = m; 4482 if (buf_mapped(bp)) { 4483 BUF_CHECK_MAPPED(bp); 4484 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4485 bp->b_pages, bp->b_npages); 4486 } else 4487 BUF_CHECK_UNMAPPED(bp); 4488 } 4489 vm_page_sunbusy(m); 4490 } 4491 vm_object_pip_wakeupn(obj, bp->b_npages); 4492 VM_OBJECT_WUNLOCK(obj); 4493 } 4494 4495 /* 4496 * vfs_page_set_valid: 4497 * 4498 * Set the valid bits in a page based on the supplied offset. The 4499 * range is restricted to the buffer's size. 4500 * 4501 * This routine is typically called after a read completes. 4502 */ 4503 static void 4504 vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4505 { 4506 vm_ooffset_t eoff; 4507 4508 /* 4509 * Compute the end offset, eoff, such that [off, eoff) does not span a 4510 * page boundary and eoff is not greater than the end of the buffer. 4511 * The end of the buffer, in this case, is our file EOF, not the 4512 * allocation size of the buffer. 4513 */ 4514 eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; 4515 if (eoff > bp->b_offset + bp->b_bcount) 4516 eoff = bp->b_offset + bp->b_bcount; 4517 4518 /* 4519 * Set valid range. This is typically the entire buffer and thus the 4520 * entire page. 4521 */ 4522 if (eoff > off) 4523 vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); 4524 } 4525 4526 /* 4527 * vfs_page_set_validclean: 4528 * 4529 * Set the valid bits and clear the dirty bits in a page based on the 4530 * supplied offset. The range is restricted to the buffer's size. 4531 */ 4532 static void 4533 vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) 4534 { 4535 vm_ooffset_t soff, eoff; 4536 4537 /* 4538 * Start and end offsets in buffer. eoff - soff may not cross a 4539 * page boundary or cross the end of the buffer. The end of the 4540 * buffer, in this case, is our file EOF, not the allocation size 4541 * of the buffer. 4542 */ 4543 soff = off; 4544 eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4545 if (eoff > bp->b_offset + bp->b_bcount) 4546 eoff = bp->b_offset + bp->b_bcount; 4547 4548 /* 4549 * Set valid range. This is typically the entire buffer and thus the 4550 * entire page. 4551 */ 4552 if (eoff > soff) { 4553 vm_page_set_validclean( 4554 m, 4555 (vm_offset_t) (soff & PAGE_MASK), 4556 (vm_offset_t) (eoff - soff) 4557 ); 4558 } 4559 } 4560 4561 /* 4562 * Ensure that all buffer pages are not exclusive busied. If any page is 4563 * exclusive busy, drain it. 4564 */ 4565 void 4566 vfs_drain_busy_pages(struct buf *bp) 4567 { 4568 vm_page_t m; 4569 int i, last_busied; 4570 4571 VM_OBJECT_ASSERT_WLOCKED(bp->b_bufobj->bo_object); 4572 last_busied = 0; 4573 for (i = 0; i < bp->b_npages; i++) { 4574 m = bp->b_pages[i]; 4575 if (vm_page_xbusied(m)) { 4576 for (; last_busied < i; last_busied++) 4577 vm_page_sbusy(bp->b_pages[last_busied]); 4578 while (vm_page_xbusied(m)) { 4579 vm_page_lock(m); 4580 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4581 vm_page_busy_sleep(m, "vbpage", true); 4582 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4583 } 4584 } 4585 } 4586 for (i = 0; i < last_busied; i++) 4587 vm_page_sunbusy(bp->b_pages[i]); 4588 } 4589 4590 /* 4591 * This routine is called before a device strategy routine. 4592 * It is used to tell the VM system that paging I/O is in 4593 * progress, and treat the pages associated with the buffer 4594 * almost as being exclusive busy. Also the object paging_in_progress 4595 * flag is handled to make sure that the object doesn't become 4596 * inconsistent. 4597 * 4598 * Since I/O has not been initiated yet, certain buffer flags 4599 * such as BIO_ERROR or B_INVAL may be in an inconsistent state 4600 * and should be ignored. 4601 */ 4602 void 4603 vfs_busy_pages(struct buf *bp, int clear_modify) 4604 { 4605 vm_object_t obj; 4606 vm_ooffset_t foff; 4607 vm_page_t m; 4608 int i; 4609 bool bogus; 4610 4611 if (!(bp->b_flags & B_VMIO)) 4612 return; 4613 4614 obj = bp->b_bufobj->bo_object; 4615 foff = bp->b_offset; 4616 KASSERT(bp->b_offset != NOOFFSET, 4617 ("vfs_busy_pages: no buffer offset")); 4618 VM_OBJECT_WLOCK(obj); 4619 vfs_drain_busy_pages(bp); 4620 if (bp->b_bufsize != 0) 4621 vfs_setdirty_locked_object(bp); 4622 bogus = false; 4623 for (i = 0; i < bp->b_npages; i++) { 4624 m = bp->b_pages[i]; 4625 4626 if ((bp->b_flags & B_CLUSTER) == 0) { 4627 vm_object_pip_add(obj, 1); 4628 vm_page_sbusy(m); 4629 } 4630 /* 4631 * When readying a buffer for a read ( i.e 4632 * clear_modify == 0 ), it is important to do 4633 * bogus_page replacement for valid pages in 4634 * partially instantiated buffers. Partially 4635 * instantiated buffers can, in turn, occur when 4636 * reconstituting a buffer from its VM backing store 4637 * base. We only have to do this if B_CACHE is 4638 * clear ( which causes the I/O to occur in the 4639 * first place ). The replacement prevents the read 4640 * I/O from overwriting potentially dirty VM-backed 4641 * pages. XXX bogus page replacement is, uh, bogus. 4642 * It may not work properly with small-block devices. 4643 * We need to find a better way. 4644 */ 4645 if (clear_modify) { 4646 pmap_remove_write(m); 4647 vfs_page_set_validclean(bp, foff, m); 4648 } else if (m->valid == VM_PAGE_BITS_ALL && 4649 (bp->b_flags & B_CACHE) == 0) { 4650 bp->b_pages[i] = bogus_page; 4651 bogus = true; 4652 } 4653 foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; 4654 } 4655 VM_OBJECT_WUNLOCK(obj); 4656 if (bogus && buf_mapped(bp)) { 4657 BUF_CHECK_MAPPED(bp); 4658 pmap_qenter(trunc_page((vm_offset_t)bp->b_data), 4659 bp->b_pages, bp->b_npages); 4660 } 4661 } 4662 4663 /* 4664 * vfs_bio_set_valid: 4665 * 4666 * Set the range within the buffer to valid. The range is 4667 * relative to the beginning of the buffer, b_offset. Note that 4668 * b_offset itself may be offset from the beginning of the first 4669 * page. 4670 */ 4671 void 4672 vfs_bio_set_valid(struct buf *bp, int base, int size) 4673 { 4674 int i, n; 4675 vm_page_t m; 4676 4677 if (!(bp->b_flags & B_VMIO)) 4678 return; 4679 4680 /* 4681 * Fixup base to be relative to beginning of first page. 4682 * Set initial n to be the maximum number of bytes in the 4683 * first page that can be validated. 4684 */ 4685 base += (bp->b_offset & PAGE_MASK); 4686 n = PAGE_SIZE - (base & PAGE_MASK); 4687 4688 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4689 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4690 m = bp->b_pages[i]; 4691 if (n > size) 4692 n = size; 4693 vm_page_set_valid_range(m, base & PAGE_MASK, n); 4694 base += n; 4695 size -= n; 4696 n = PAGE_SIZE; 4697 } 4698 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4699 } 4700 4701 /* 4702 * vfs_bio_clrbuf: 4703 * 4704 * If the specified buffer is a non-VMIO buffer, clear the entire 4705 * buffer. If the specified buffer is a VMIO buffer, clear and 4706 * validate only the previously invalid portions of the buffer. 4707 * This routine essentially fakes an I/O, so we need to clear 4708 * BIO_ERROR and B_INVAL. 4709 * 4710 * Note that while we only theoretically need to clear through b_bcount, 4711 * we go ahead and clear through b_bufsize. 4712 */ 4713 void 4714 vfs_bio_clrbuf(struct buf *bp) 4715 { 4716 int i, j, mask, sa, ea, slide; 4717 4718 if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { 4719 clrbuf(bp); 4720 return; 4721 } 4722 bp->b_flags &= ~B_INVAL; 4723 bp->b_ioflags &= ~BIO_ERROR; 4724 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object); 4725 if ((bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE) && 4726 (bp->b_offset & PAGE_MASK) == 0) { 4727 if (bp->b_pages[0] == bogus_page) 4728 goto unlock; 4729 mask = (1 << (bp->b_bufsize / DEV_BSIZE)) - 1; 4730 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[0]->object); 4731 if ((bp->b_pages[0]->valid & mask) == mask) 4732 goto unlock; 4733 if ((bp->b_pages[0]->valid & mask) == 0) { 4734 pmap_zero_page_area(bp->b_pages[0], 0, bp->b_bufsize); 4735 bp->b_pages[0]->valid |= mask; 4736 goto unlock; 4737 } 4738 } 4739 sa = bp->b_offset & PAGE_MASK; 4740 slide = 0; 4741 for (i = 0; i < bp->b_npages; i++, sa = 0) { 4742 slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); 4743 ea = slide & PAGE_MASK; 4744 if (ea == 0) 4745 ea = PAGE_SIZE; 4746 if (bp->b_pages[i] == bogus_page) 4747 continue; 4748 j = sa / DEV_BSIZE; 4749 mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; 4750 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[i]->object); 4751 if ((bp->b_pages[i]->valid & mask) == mask) 4752 continue; 4753 if ((bp->b_pages[i]->valid & mask) == 0) 4754 pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); 4755 else { 4756 for (; sa < ea; sa += DEV_BSIZE, j++) { 4757 if ((bp->b_pages[i]->valid & (1 << j)) == 0) { 4758 pmap_zero_page_area(bp->b_pages[i], 4759 sa, DEV_BSIZE); 4760 } 4761 } 4762 } 4763 bp->b_pages[i]->valid |= mask; 4764 } 4765 unlock: 4766 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object); 4767 bp->b_resid = 0; 4768 } 4769 4770 void 4771 vfs_bio_bzero_buf(struct buf *bp, int base, int size) 4772 { 4773 vm_page_t m; 4774 int i, n; 4775 4776 if (buf_mapped(bp)) { 4777 BUF_CHECK_MAPPED(bp); 4778 bzero(bp->b_data + base, size); 4779 } else { 4780 BUF_CHECK_UNMAPPED(bp); 4781 n = PAGE_SIZE - (base & PAGE_MASK); 4782 for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { 4783 m = bp->b_pages[i]; 4784 if (n > size) 4785 n = size; 4786 pmap_zero_page_area(m, base & PAGE_MASK, n); 4787 base += n; 4788 size -= n; 4789 n = PAGE_SIZE; 4790 } 4791 } 4792 } 4793 4794 /* 4795 * Update buffer flags based on I/O request parameters, optionally releasing the 4796 * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, 4797 * where they may be placed on a page queue (VMIO) or freed immediately (direct 4798 * I/O). Otherwise the buffer is released to the cache. 4799 */ 4800 static void 4801 b_io_dismiss(struct buf *bp, int ioflag, bool release) 4802 { 4803 4804 KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, 4805 ("buf %p non-VMIO noreuse", bp)); 4806 4807 if ((ioflag & IO_DIRECT) != 0) 4808 bp->b_flags |= B_DIRECT; 4809 if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { 4810 bp->b_flags |= B_RELBUF; 4811 if ((ioflag & IO_NOREUSE) != 0) 4812 bp->b_flags |= B_NOREUSE; 4813 if (release) 4814 brelse(bp); 4815 } else if (release) 4816 bqrelse(bp); 4817 } 4818 4819 void 4820 vfs_bio_brelse(struct buf *bp, int ioflag) 4821 { 4822 4823 b_io_dismiss(bp, ioflag, true); 4824 } 4825 4826 void 4827 vfs_bio_set_flags(struct buf *bp, int ioflag) 4828 { 4829 4830 b_io_dismiss(bp, ioflag, false); 4831 } 4832 4833 /* 4834 * vm_hold_load_pages and vm_hold_free_pages get pages into 4835 * a buffers address space. The pages are anonymous and are 4836 * not associated with a file object. 4837 */ 4838 static void 4839 vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) 4840 { 4841 vm_offset_t pg; 4842 vm_page_t p; 4843 int index; 4844 4845 BUF_CHECK_MAPPED(bp); 4846 4847 to = round_page(to); 4848 from = round_page(from); 4849 index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4850 4851 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 4852 /* 4853 * note: must allocate system pages since blocking here 4854 * could interfere with paging I/O, no matter which 4855 * process we are. 4856 */ 4857 p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | 4858 VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | 4859 VM_ALLOC_WAITOK); 4860 pmap_qenter(pg, &p, 1); 4861 bp->b_pages[index] = p; 4862 } 4863 bp->b_npages = index; 4864 } 4865 4866 /* Return pages associated with this buf to the vm system */ 4867 static void 4868 vm_hold_free_pages(struct buf *bp, int newbsize) 4869 { 4870 vm_offset_t from; 4871 vm_page_t p; 4872 int index, newnpages; 4873 4874 BUF_CHECK_MAPPED(bp); 4875 4876 from = round_page((vm_offset_t)bp->b_data + newbsize); 4877 newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; 4878 if (bp->b_npages > newnpages) 4879 pmap_qremove(from, bp->b_npages - newnpages); 4880 for (index = newnpages; index < bp->b_npages; index++) { 4881 p = bp->b_pages[index]; 4882 bp->b_pages[index] = NULL; 4883 p->wire_count--; 4884 vm_page_free(p); 4885 } 4886 vm_wire_sub(bp->b_npages - newnpages); 4887 bp->b_npages = newnpages; 4888 } 4889 4890 /* 4891 * Map an IO request into kernel virtual address space. 4892 * 4893 * All requests are (re)mapped into kernel VA space. 4894 * Notice that we use b_bufsize for the size of the buffer 4895 * to be mapped. b_bcount might be modified by the driver. 4896 * 4897 * Note that even if the caller determines that the address space should 4898 * be valid, a race or a smaller-file mapped into a larger space may 4899 * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST 4900 * check the return value. 4901 * 4902 * This function only works with pager buffers. 4903 */ 4904 int 4905 vmapbuf(struct buf *bp, int mapbuf) 4906 { 4907 vm_prot_t prot; 4908 int pidx; 4909 4910 if (bp->b_bufsize < 0) 4911 return (-1); 4912 prot = VM_PROT_READ; 4913 if (bp->b_iocmd == BIO_READ) 4914 prot |= VM_PROT_WRITE; /* Less backwards than it looks */ 4915 if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, 4916 (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, 4917 btoc(MAXPHYS))) < 0) 4918 return (-1); 4919 bp->b_npages = pidx; 4920 bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; 4921 if (mapbuf || !unmapped_buf_allowed) { 4922 pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); 4923 bp->b_data = bp->b_kvabase + bp->b_offset; 4924 } else 4925 bp->b_data = unmapped_buf; 4926 return(0); 4927 } 4928 4929 /* 4930 * Free the io map PTEs associated with this IO operation. 4931 * We also invalidate the TLB entries and restore the original b_addr. 4932 * 4933 * This function only works with pager buffers. 4934 */ 4935 void 4936 vunmapbuf(struct buf *bp) 4937 { 4938 int npages; 4939 4940 npages = bp->b_npages; 4941 if (buf_mapped(bp)) 4942 pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); 4943 vm_page_unhold_pages(bp->b_pages, npages); 4944 4945 bp->b_data = unmapped_buf; 4946 } 4947 4948 void 4949 bdone(struct buf *bp) 4950 { 4951 struct mtx *mtxp; 4952 4953 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4954 mtx_lock(mtxp); 4955 bp->b_flags |= B_DONE; 4956 wakeup(bp); 4957 mtx_unlock(mtxp); 4958 } 4959 4960 void 4961 bwait(struct buf *bp, u_char pri, const char *wchan) 4962 { 4963 struct mtx *mtxp; 4964 4965 mtxp = mtx_pool_find(mtxpool_sleep, bp); 4966 mtx_lock(mtxp); 4967 while ((bp->b_flags & B_DONE) == 0) 4968 msleep(bp, mtxp, pri, wchan, 0); 4969 mtx_unlock(mtxp); 4970 } 4971 4972 int 4973 bufsync(struct bufobj *bo, int waitfor) 4974 { 4975 4976 return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); 4977 } 4978 4979 void 4980 bufstrategy(struct bufobj *bo, struct buf *bp) 4981 { 4982 int i = 0; 4983 struct vnode *vp; 4984 4985 vp = bp->b_vp; 4986 KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); 4987 KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, 4988 ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); 4989 i = VOP_STRATEGY(vp, bp); 4990 KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); 4991 } 4992 4993 /* 4994 * Initialize a struct bufobj before use. Memory is assumed zero filled. 4995 */ 4996 void 4997 bufobj_init(struct bufobj *bo, void *private) 4998 { 4999 static volatile int bufobj_cleanq; 5000 5001 bo->bo_domain = 5002 atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; 5003 rw_init(BO_LOCKPTR(bo), "bufobj interlock"); 5004 bo->bo_private = private; 5005 TAILQ_INIT(&bo->bo_clean.bv_hd); 5006 TAILQ_INIT(&bo->bo_dirty.bv_hd); 5007 } 5008 5009 void 5010 bufobj_wrefl(struct bufobj *bo) 5011 { 5012 5013 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5014 ASSERT_BO_WLOCKED(bo); 5015 bo->bo_numoutput++; 5016 } 5017 5018 void 5019 bufobj_wref(struct bufobj *bo) 5020 { 5021 5022 KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); 5023 BO_LOCK(bo); 5024 bo->bo_numoutput++; 5025 BO_UNLOCK(bo); 5026 } 5027 5028 void 5029 bufobj_wdrop(struct bufobj *bo) 5030 { 5031 5032 KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); 5033 BO_LOCK(bo); 5034 KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); 5035 if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { 5036 bo->bo_flag &= ~BO_WWAIT; 5037 wakeup(&bo->bo_numoutput); 5038 } 5039 BO_UNLOCK(bo); 5040 } 5041 5042 int 5043 bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) 5044 { 5045 int error; 5046 5047 KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); 5048 ASSERT_BO_WLOCKED(bo); 5049 error = 0; 5050 while (bo->bo_numoutput) { 5051 bo->bo_flag |= BO_WWAIT; 5052 error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), 5053 slpflag | (PRIBIO + 1), "bo_wwait", timeo); 5054 if (error) 5055 break; 5056 } 5057 return (error); 5058 } 5059 5060 /* 5061 * Set bio_data or bio_ma for struct bio from the struct buf. 5062 */ 5063 void 5064 bdata2bio(struct buf *bp, struct bio *bip) 5065 { 5066 5067 if (!buf_mapped(bp)) { 5068 KASSERT(unmapped_buf_allowed, ("unmapped")); 5069 bip->bio_ma = bp->b_pages; 5070 bip->bio_ma_n = bp->b_npages; 5071 bip->bio_data = unmapped_buf; 5072 bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; 5073 bip->bio_flags |= BIO_UNMAPPED; 5074 KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / 5075 PAGE_SIZE == bp->b_npages, 5076 ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, 5077 (long long)bip->bio_length, bip->bio_ma_n)); 5078 } else { 5079 bip->bio_data = bp->b_data; 5080 bip->bio_ma = NULL; 5081 } 5082 } 5083 5084 /* 5085 * The MIPS pmap code currently doesn't handle aliased pages. 5086 * The VIPT caches may not handle page aliasing themselves, leading 5087 * to data corruption. 5088 * 5089 * As such, this code makes a system extremely unhappy if said 5090 * system doesn't support unaliasing the above situation in hardware. 5091 * Some "recent" systems (eg some mips24k/mips74k cores) don't enable 5092 * this feature at build time, so it has to be handled in software. 5093 * 5094 * Once the MIPS pmap/cache code grows to support this function on 5095 * earlier chips, it should be flipped back off. 5096 */ 5097 #ifdef __mips__ 5098 static int buf_pager_relbuf = 1; 5099 #else 5100 static int buf_pager_relbuf = 0; 5101 #endif 5102 SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, 5103 &buf_pager_relbuf, 0, 5104 "Make buffer pager release buffers after reading"); 5105 5106 /* 5107 * The buffer pager. It uses buffer reads to validate pages. 5108 * 5109 * In contrast to the generic local pager from vm/vnode_pager.c, this 5110 * pager correctly and easily handles volumes where the underlying 5111 * device block size is greater than the machine page size. The 5112 * buffer cache transparently extends the requested page run to be 5113 * aligned at the block boundary, and does the necessary bogus page 5114 * replacements in the addends to avoid obliterating already valid 5115 * pages. 5116 * 5117 * The only non-trivial issue is that the exclusive busy state for 5118 * pages, which is assumed by the vm_pager_getpages() interface, is 5119 * incompatible with the VMIO buffer cache's desire to share-busy the 5120 * pages. This function performs a trivial downgrade of the pages' 5121 * state before reading buffers, and a less trivial upgrade from the 5122 * shared-busy to excl-busy state after the read. 5123 */ 5124 int 5125 vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, 5126 int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, 5127 vbg_get_blksize_t get_blksize) 5128 { 5129 vm_page_t m; 5130 vm_object_t object; 5131 struct buf *bp; 5132 struct mount *mp; 5133 daddr_t lbn, lbnp; 5134 vm_ooffset_t la, lb, poff, poffe; 5135 long bsize; 5136 int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b; 5137 bool redo, lpart; 5138 5139 object = vp->v_object; 5140 mp = vp->v_mount; 5141 error = 0; 5142 la = IDX_TO_OFF(ma[count - 1]->pindex); 5143 if (la >= object->un_pager.vnp.vnp_size) 5144 return (VM_PAGER_BAD); 5145 5146 /* 5147 * Change the meaning of la from where the last requested page starts 5148 * to where it ends, because that's the end of the requested region 5149 * and the start of the potential read-ahead region. 5150 */ 5151 la += PAGE_SIZE; 5152 lpart = la > object->un_pager.vnp.vnp_size; 5153 bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex))); 5154 5155 /* 5156 * Calculate read-ahead, behind and total pages. 5157 */ 5158 pgsin = count; 5159 lb = IDX_TO_OFF(ma[0]->pindex); 5160 pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); 5161 pgsin += pgsin_b; 5162 if (rbehind != NULL) 5163 *rbehind = pgsin_b; 5164 pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); 5165 if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) 5166 pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, 5167 PAGE_SIZE) - la); 5168 pgsin += pgsin_a; 5169 if (rahead != NULL) 5170 *rahead = pgsin_a; 5171 VM_CNT_INC(v_vnodein); 5172 VM_CNT_ADD(v_vnodepgsin, pgsin); 5173 5174 br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) 5175 != 0) ? GB_UNMAPPED : 0; 5176 VM_OBJECT_WLOCK(object); 5177 again: 5178 for (i = 0; i < count; i++) 5179 vm_page_busy_downgrade(ma[i]); 5180 VM_OBJECT_WUNLOCK(object); 5181 5182 lbnp = -1; 5183 for (i = 0; i < count; i++) { 5184 m = ma[i]; 5185 5186 /* 5187 * Pages are shared busy and the object lock is not 5188 * owned, which together allow for the pages' 5189 * invalidation. The racy test for validity avoids 5190 * useless creation of the buffer for the most typical 5191 * case when invalidation is not used in redo or for 5192 * parallel read. The shared->excl upgrade loop at 5193 * the end of the function catches the race in a 5194 * reliable way (protected by the object lock). 5195 */ 5196 if (m->valid == VM_PAGE_BITS_ALL) 5197 continue; 5198 5199 poff = IDX_TO_OFF(m->pindex); 5200 poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); 5201 for (; poff < poffe; poff += bsize) { 5202 lbn = get_lblkno(vp, poff); 5203 if (lbn == lbnp) 5204 goto next_page; 5205 lbnp = lbn; 5206 5207 bsize = get_blksize(vp, lbn); 5208 error = bread_gb(vp, lbn, bsize, curthread->td_ucred, 5209 br_flags, &bp); 5210 if (error != 0) 5211 goto end_pages; 5212 if (LIST_EMPTY(&bp->b_dep)) { 5213 /* 5214 * Invalidation clears m->valid, but 5215 * may leave B_CACHE flag if the 5216 * buffer existed at the invalidation 5217 * time. In this case, recycle the 5218 * buffer to do real read on next 5219 * bread() after redo. 5220 * 5221 * Otherwise B_RELBUF is not strictly 5222 * necessary, enable to reduce buf 5223 * cache pressure. 5224 */ 5225 if (buf_pager_relbuf || 5226 m->valid != VM_PAGE_BITS_ALL) 5227 bp->b_flags |= B_RELBUF; 5228 5229 bp->b_flags &= ~B_NOCACHE; 5230 brelse(bp); 5231 } else { 5232 bqrelse(bp); 5233 } 5234 } 5235 KASSERT(1 /* racy, enable for debugging */ || 5236 m->valid == VM_PAGE_BITS_ALL || i == count - 1, 5237 ("buf %d %p invalid", i, m)); 5238 if (i == count - 1 && lpart) { 5239 VM_OBJECT_WLOCK(object); 5240 if (m->valid != 0 && 5241 m->valid != VM_PAGE_BITS_ALL) 5242 vm_page_zero_invalid(m, TRUE); 5243 VM_OBJECT_WUNLOCK(object); 5244 } 5245 next_page:; 5246 } 5247 end_pages: 5248 5249 VM_OBJECT_WLOCK(object); 5250 redo = false; 5251 for (i = 0; i < count; i++) { 5252 vm_page_sunbusy(ma[i]); 5253 ma[i] = vm_page_grab(object, ma[i]->pindex, VM_ALLOC_NORMAL); 5254 5255 /* 5256 * Since the pages were only sbusy while neither the 5257 * buffer nor the object lock was held by us, or 5258 * reallocated while vm_page_grab() slept for busy 5259 * relinguish, they could have been invalidated. 5260 * Recheck the valid bits and re-read as needed. 5261 * 5262 * Note that the last page is made fully valid in the 5263 * read loop, and partial validity for the page at 5264 * index count - 1 could mean that the page was 5265 * invalidated or removed, so we must restart for 5266 * safety as well. 5267 */ 5268 if (ma[i]->valid != VM_PAGE_BITS_ALL) 5269 redo = true; 5270 } 5271 if (redo && error == 0) 5272 goto again; 5273 VM_OBJECT_WUNLOCK(object); 5274 return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); 5275 } 5276 5277 #include "opt_ddb.h" 5278 #ifdef DDB 5279 #include <ddb/ddb.h> 5280 5281 /* DDB command to show buffer data */ 5282 DB_SHOW_COMMAND(buffer, db_show_buffer) 5283 { 5284 /* get args */ 5285 struct buf *bp = (struct buf *)addr; 5286 #ifdef FULL_BUF_TRACKING 5287 uint32_t i, j; 5288 #endif 5289 5290 if (!have_addr) { 5291 db_printf("usage: show buffer <addr>\n"); 5292 return; 5293 } 5294 5295 db_printf("buf at %p\n", bp); 5296 db_printf("b_flags = 0x%b, b_xflags=0x%b, b_vflags=0x%b\n", 5297 (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, 5298 PRINT_BUF_XFLAGS, (u_int)bp->b_vflags, PRINT_BUF_VFLAGS); 5299 db_printf( 5300 "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" 5301 "b_bufobj = (%p), b_data = %p, b_blkno = %jd, b_lblkno = %jd, " 5302 "b_dep = %p\n", 5303 bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, 5304 bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, 5305 (intmax_t)bp->b_lblkno, bp->b_dep.lh_first); 5306 db_printf("b_kvabase = %p, b_kvasize = %d\n", 5307 bp->b_kvabase, bp->b_kvasize); 5308 if (bp->b_npages) { 5309 int i; 5310 db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); 5311 for (i = 0; i < bp->b_npages; i++) { 5312 vm_page_t m; 5313 m = bp->b_pages[i]; 5314 if (m != NULL) 5315 db_printf("(%p, 0x%lx, 0x%lx)", m->object, 5316 (u_long)m->pindex, 5317 (u_long)VM_PAGE_TO_PHYS(m)); 5318 else 5319 db_printf("( ??? )"); 5320 if ((i + 1) < bp->b_npages) 5321 db_printf(","); 5322 } 5323 db_printf("\n"); 5324 } 5325 BUF_LOCKPRINTINFO(bp); 5326 #if defined(FULL_BUF_TRACKING) 5327 db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); 5328 5329 i = bp->b_io_tcnt % BUF_TRACKING_SIZE; 5330 for (j = 1; j <= BUF_TRACKING_SIZE; j++) { 5331 if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) 5332 continue; 5333 db_printf(" %2u: %s\n", j, 5334 bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); 5335 } 5336 #elif defined(BUF_TRACKING) 5337 db_printf("b_io_tracking: %s\n", bp->b_io_tracking); 5338 #endif 5339 db_printf(" "); 5340 } 5341 5342 DB_SHOW_COMMAND(bufqueues, bufqueues) 5343 { 5344 struct bufdomain *bd; 5345 struct buf *bp; 5346 long total; 5347 int i, j, cnt; 5348 5349 db_printf("bqempty: %d\n", bqempty.bq_len); 5350 5351 for (i = 0; i < buf_domains; i++) { 5352 bd = &bdomain[i]; 5353 db_printf("Buf domain %d\n", i); 5354 db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); 5355 db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); 5356 db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); 5357 db_printf("\n"); 5358 db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); 5359 db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); 5360 db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); 5361 db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); 5362 db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); 5363 db_printf("\n"); 5364 db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); 5365 db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); 5366 db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); 5367 db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); 5368 db_printf("\n"); 5369 total = 0; 5370 TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) 5371 total += bp->b_bufsize; 5372 db_printf("\tcleanq count\t%d (%ld)\n", 5373 bd->bd_cleanq->bq_len, total); 5374 total = 0; 5375 TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) 5376 total += bp->b_bufsize; 5377 db_printf("\tdirtyq count\t%d (%ld)\n", 5378 bd->bd_dirtyq.bq_len, total); 5379 db_printf("\twakeup\t\t%d\n", bd->bd_wanted); 5380 db_printf("\tlim\t\t%d\n", bd->bd_lim); 5381 db_printf("\tCPU "); 5382 for (j = 0; j <= mp_maxid; j++) 5383 db_printf("%d, ", bd->bd_subq[j].bq_len); 5384 db_printf("\n"); 5385 cnt = 0; 5386 total = 0; 5387 for (j = 0; j < nbuf; j++) 5388 if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) { 5389 cnt++; 5390 total += buf[j].b_bufsize; 5391 } 5392 db_printf("\tLocked buffers: %d space %ld\n", cnt, total); 5393 cnt = 0; 5394 total = 0; 5395 for (j = 0; j < nbuf; j++) 5396 if (buf[j].b_domain == i) { 5397 cnt++; 5398 total += buf[j].b_bufsize; 5399 } 5400 db_printf("\tTotal buffers: %d space %ld\n", cnt, total); 5401 } 5402 } 5403 5404 DB_SHOW_COMMAND(lockedbufs, lockedbufs) 5405 { 5406 struct buf *bp; 5407 int i; 5408 5409 for (i = 0; i < nbuf; i++) { 5410 bp = &buf[i]; 5411 if (BUF_ISLOCKED(bp)) { 5412 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5413 db_printf("\n"); 5414 if (db_pager_quit) 5415 break; 5416 } 5417 } 5418 } 5419 5420 DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) 5421 { 5422 struct vnode *vp; 5423 struct buf *bp; 5424 5425 if (!have_addr) { 5426 db_printf("usage: show vnodebufs <addr>\n"); 5427 return; 5428 } 5429 vp = (struct vnode *)addr; 5430 db_printf("Clean buffers:\n"); 5431 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { 5432 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5433 db_printf("\n"); 5434 } 5435 db_printf("Dirty buffers:\n"); 5436 TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { 5437 db_show_buffer((uintptr_t)bp, 1, 0, NULL); 5438 db_printf("\n"); 5439 } 5440 } 5441 5442 DB_COMMAND(countfreebufs, db_coundfreebufs) 5443 { 5444 struct buf *bp; 5445 int i, used = 0, nfree = 0; 5446 5447 if (have_addr) { 5448 db_printf("usage: countfreebufs\n"); 5449 return; 5450 } 5451 5452 for (i = 0; i < nbuf; i++) { 5453 bp = &buf[i]; 5454 if (bp->b_qindex == QUEUE_EMPTY) 5455 nfree++; 5456 else 5457 used++; 5458 } 5459 5460 db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, 5461 nfree + used); 5462 db_printf("numfreebuffers is %d\n", numfreebuffers); 5463 } 5464 #endif /* DDB */ 5465